APF-YOLOV8: Enhancing Multiscale Detection and Intra-Class Variance Handling for UAV-Based Insulator Power Line Inspections

Multi-scale feature management

Effectively managing multi-scale features is a cornerstone of robust object detection, enabling models to handle objects of varying sizes and resolutions. YOLO-based models have tackled this challenge through diverse architectural innovations and feature integration strategies.

The YOLOv8-baseline model¹⁴ lays the groundwork by incorporating a Pyramid Attention Network-Feature Pyramid Network (PAN-FPN),^15,16 which facilitates the fusion of high-level semantic features with low-level spatial details. This combination ensures robust detection of objects at multiple scales. Complementing this, the integration of the Cross-Stage Partial Network (CSPNet)¹⁷ optimizes feature reuse, making multi-scale feature extraction both efficient and effective.

Building on this solid foundation, YOLOv8-P2 introduces an additional P2 layer that significantly enhances granularity, particularly for small-object detection. This improvement proves invaluable in environments with dynamic object scales, such as those encountered in UAV-based applications. Further extending this lineage, YOLOv9¹⁸ incorporates the Generalized Efficient Layer Aggregation Network (GELAN), an advanced architecture designed to minimize information loss during forward propagation. GELAN’s ability to aggregate features without introducing bottlenecks allows for precise detection of both small and large objects, a capability further enhanced by the use of Programmable Gradient Information (PGI) to ensure robust training.

While these models focus on general-purpose enhancements, other YOLO variants target specific challenges in multi-scale detection. For instance, YOLOv8-MAFPN¹⁹ introduces the Multi-Auxiliary Feature Pyramid Network, which leverages a dual fusion strategy: the Superficial Assisted Fusion (SAF) module integrates shallow features for preserving small-object details, while the Advanced Assisted Fusion (AAF) module conveys diverse gradient information for deeper feature learning. Similarly, YOLOv8-HSFPN²⁰ employs a High-level Screening Feature Pyramid (HS-FPN), dynamically filtering and fusing features to achieve a balanced representation. This approach is particularly effective in complex scenarios like medical imaging and industrial inspection.

Another noteworthy model, YOLOv8-bifpn,²¹ incorporates a Bi-Directional Feature Pyramid Network (BiFPN), which allows for seamless top-down and bottom-up feature flow. This bidirectional connectivity, coupled with weighted fusion, enables efficient handling of multi-scale features, making it suitable for applications requiring both speed and accuracy.

Attention mechanisms also play a critical role in enhancing multi-scale capabilities. For example, YOLOv8-ASF-P2²² integrates Attentional Scale Sequence Fusion, which prioritizes features based on their relevance at different scales. Initially developed for medical imaging, this method has shown promise in detecting small and dense objects in other challenging domains, such as aerial imagery. Likewise, YOLOv8-CFPT-P2345²³ introduces the Cross-Layer Feature Pyramid Transformer (CFPT), which avoids traditional semantic gaps by leveraging attention blocks for cross-layer global information fusion. This design is particularly advantageous for detecting small objects where detail preservation is critical.

Finally, YOLOv10²⁴ combines these advancements into a holistic framework. Its efficiency-accuracy-driven design incorporates lightweight classification heads, spatial-channel decoupled downsampling, and rank-guided blocks. These features collectively enable high-performance multi-scale detection with reduced computational overhead, making it ideal for real-world, resource-constrained applications.

Intra-class variance handling

Addressing intra-class variance—where objects of the same class exhibit significant differences in appearance due to material, texture, or environmental conditions—is another critical challenge in object detection. YOLO-based models have approached this problem with a variety of innovative techniques.

The YOLOv8-baseline¹⁴ model addresses intra-class variance through its anchor-free design, which allows for greater flexibility in adapting to variations in object shape and size. Building on this, YOLOv8-P2 leverages its P2 layer to extract fine-grained features, making it particularly adept at differentiating visually similar objects within the same class.

A more advanced solution comes from YOLOv9, which introduces Programmable Gradient Information (PGI) to focus the model’s attention on distinguishing features. By ensuring reliable gradient flow and minimizing information loss, PGI enables robust handling of intra-class variance, even in dynamic environments. Complementing this, the GELAN architecture efficiently aggregates features while preserving subtle details, enhancing the model’s ability to differentiate between closely related objects.

For scenarios requiring a more nuanced approach, YOLOv8-MAFPN aligns features across scales to capture subtle variations within the same class. Similarly, YOLOv8-HSFPN combines shallow and deep feature extraction through its HS-FPN module, achieving better representation in domains like medical imaging, where intra-class variance is particularly pronounced.

Other YOLO variants incorporate specialized mechanisms to tackle this issue. YOLOv8-goldyolo,²⁵ for example, employs a Gather-and-Distribute (GD) mechanism that aggregates features to enhance differentiation. It further introduces MAE-style pretraining, an innovative approach that leverages unsupervised learning to improve robustness. Likewise, YOLOv8-ASF-P2 integrates Channel and Position Attention Mechanisms (CPAM) to emphasize the most discriminative features, demonstrating superior performance in detecting small and densely packed objects, such as in cell segmentation.

Finally, YOLOv10 takes a step forward with its consistent dual assignments for NMS-free training. By aligning supervisory signals, this approach improves feature extraction and strengthens the model’s ability to handle intra-class variance across diverse conditions. Coupled with its efficiency-driven design, YOLOv10 delivers robust performance while maintaining low computational costs, cementing its place as a cutting-edge solution for object detection.

Our contribution

In this landscape, APF-YOLO introduces a novel neck configuration to address key challenges in multi-scale feature management and intra-class variance handling, particularly for UAV-based power line inspections. The proposed Adaptive Path Fusion (APF) neck synthesizes advancements from previous YOLO-based models, offering seamless integration of multi-scale features while maintaining computational efficiency. Additionally, the Adaptive Feature Alignment Module (AFAM) ensures robust alignment of features across scales, enabling accurate detection of objects with diverse sizes and characteristics. APF-YOLO integrates advanced attention mechanisms, including Efficient Local Attention (ELA), Efficient Channel Attention (ECA), and Context Anchor Attention (CAA). These mechanisms amplify the model’s ability to detect and differentiate objects under varying environmental and geometric conditions. Collectively, these innovations position APF-YOLO as a significant advancement in UAV-based detection, particularly in addressing the nuanced challenges of insulator detection.

To provide a comprehensive overview of the advancements in YOLO-based neck configurations and their approaches to addressing multi-scale object detection and intra-class variance handling, Table 1 summarizes the key contributions and release years of the reviewed models. This table highlights the evolutionary trajectory of these architectures and situates APF-YOLO within the broader landscape of YOLO-based innovations.

A. Problem statement

Modern YOLO models adopt a three-component architecture comprising a backbone, neck, and head. While the backbone extracts spatial and semantic features and the head generates predictions, the neck is pivotal in bridging these two components. By fusing multiscale features, the neck ensures the detection of objects of varying sizes and appearances, making it a critical area for improvement in complex detection tasks.

In the context of UAV-based power line inspections, detecting insulators presents unique challenges due to significant intra-class variance. Insulators—composed of materials like glass, composite, and porcelain—share a similar global shape but differ subtly in material texture and appearance. This poses a dual challenge: the model must preserve high-level semantic information to recognize shared global shapes while capturing fine-grained spatial details to differentiate between types.

The neck plays a vital role in this process by managing multiscale feature fusion. For insulators, preserving semantic consistency during feature fusion is essential to handle intra-class variance effectively. The global shape must be retained to ensure the model recognizes insulators as a single class, as insulators generally share a similar global shape regardless of their type. Standard architectures like PAN-FPN offer a partial solution to these challenges but leave room for optimization, as they often struggle to achieve a perfect balance between fine-grained details and high-level semantic features.

To resume, by focusing on neck improvements (not backbone or head), our approach seeks to robustly address multi-scale and intra-class variance challenges, ensuring accurate and reliable insulator detection in UAV-based power line inspections.

B. Introduction to YOLOv8

YOLOv8¹⁴ represents a significant milestone in the evolution of the YOLO series, renowned for its exceptional performance in real-time object detection. Real-time detection refers to a model’s ability to swiftly and accurately identify and localize objects with minimal latency, a capability crucial for applications like autonomous driving, live video tracking, and UAV-based power line inspections YOLOv8’s balance of speed, accuracy, and efficiency makes it particularly well-suited for these time-sensitive tasks.²⁶

One of YOLOv8’s key strengths lies in its architecture, optimized for resource-constrained environments. It offers lightweight versions, such as nano and small-sized models, which achieve high performance even on devices with limited computational capacity. This adaptability is especially advantageous in UAV-based power line inspections, where computational efficiency and portability are essential for real-time anomaly detection.

YOLOv8 introduces several architectural innovations that enhance object detection performance. Its anchor-free head architecture decouples classification and localization, enabling the model to independently optimize these tasks. This decoupling improves both the accuracy of object identification and the precision of bounding box predictions. Advanced loss functions, such as CIoU Loss for bounding box regression,²⁷ DFL (Distribution Focal Loss)²⁸ for improved classification, and VFL (Varifocal Loss)²⁹ for addressing class imbalance, further refine the model’s detection capabilities.

The introduction of the C2f module marks another major improvement over its predecessor, the CSP (Cross-Stage Partial) layer in YOLOv5. By enabling more efficient feature integration and better context understanding, the C2f module enhances YOLOv8’s ability to manage multiscale detection. This is particularly relevant for detecting objects like insulators, which vary in size and appearance but share similar global shapes, as in UAV inspection tasks.

Another notable advantage of YOLOv8 is its active development within the open-source community.³⁰ Regular updates and contributions ensure that the model evolves alongside advancements in object detection, maintaining its position at the forefront of the field. This ongoing support also facilitates its adaptability to various applications, from industrial monitoring to UAV operations.

In conclusion, YOLOv8 serves as an excellent baseline model for real-time object detection tasks, offering a robust combination of speed, accuracy, and efficiency. These strengths underscore YOLOv8’s suitability for our study, where robust detection performance is crucial in managing multiscale objects and subtle material differences in UAV-based power line inspections.

C. The architecture of YOLOv8

The YOLOv8 architecture integrates a Path Aggregation Network combined with a Feature Pyramid Network (PAN-FPN) as its neck component, crucial for addressing multiscale object detection challenges. The PAN-FPN establishes a connection between the backbone, responsible for extracting feature maps, and the detection head, which generates the final predictions. Its primary function is to fuse multiscale features, ensuring that the model can accurately detect objects of various sizes by combining both high-resolution and low-resolution information.

This section will provide an explanation of the PAN-FPN neck, supported by Figure 2, which illustrates the feature propagation from the backbone to the detection head. We begin with an overview of the backbone and detection head, followed by a detailed discussion of the neck, which serves as the primary focus of this section due to its pivotal role in improving multiscale/intra class variance object detection.

Figure 2. Diagrammatic representation of YOLOv8 network architecture.

Backbone Overview

The backbone in YOLOv8 is responsible for extracting hierarchical features from the input image at various levels of abstraction, producing feature maps at multiple resolutions. Lower layers of the backbone focus on spatially detailed information (high resolution), whereas deeper layers capture semantic-rich, low-resolution features. However, these feature maps need to be combined effectively for robust object detection across multiple scales. This is where the PAN-FPN neck plays a critical role, merging these multiscale features to ensure that both small and large objects are detected with high precision. In the context of Figure 2, the backbone is represented on the left, with feature maps being processed at various stages (P3, P4, P5) before entering the neck component.

The Role of PAN-FPN in the Neck

The PAN-FPN neck introduces a pyramidal structure that facilitates both bottom-up and top-down feature propagation, as seen in the central portion of Figure 2. This bidirectional flow is designed to capture fine-grained spatial details while maintaining high-level semantic context, which is essential for detecting objects of different sizes in complex scenes.

Bottom-Up Pathway (From P5 to P3)

The bottom-up pathway begins with P5, the deepest feature map output by the backbone, characterized by low spatial resolution but rich semantic information. This pathway progressively refines spatial detail and propagates semantic richness upward:

Upsampling: The Upsample Block (e.g., Block 10 in Figure 2) performs a 2x upsampling of the P5 feature map to match the resolution of the P4 map. This operation enlarges the spatial dimensions while preserving semantic information.

Concatenation: The upsampled P5 features are concatenated with the P4 feature map via a Concat Block (Block 11). This fusion integrates high-resolution spatial information from P4 with the semantic depth of P5, enhancing the network’s ability to handle multiscale objects.

C2f Block Refinement: After concatenation, the combined features pass through a C2f Block (Block 12). This block applies lightweight convolutions and shortcut connections to refine the fused feature map, balancing computational efficiency with feature enhancement.

SPP Module: At the P5 stage, the SPP (Spatial Pyramid Pooling) Module (Block 9) pools features at multiple scales to expand the receptive field, enabling the network to detect large objects requiring broader contextual understanding.

Top of the Pyramid (P3)

At the top of the pyramid, the network processes P3, the highest-resolution feature map. This stage prioritizes fine-grained spatial detail for detecting smaller objects:

Upsampled Features: The upsampled P4 features (processed through Block 10) are fused with the high-resolution P3 features via a Concat Block (Block 17). This operation ensures that both small-scale spatial details and intermediate semantic information are preserved.

Semantic Enrichment: The fused features pass through another C2f Block (Block 18) to enhance feature representation while maintaining a lightweight computational footprint. This integration ensures that small objects benefit from both spatial precision and the semantic depth retained from deeper layers.

Top-Down Pathway (From P3 to P5)

The top-down pathway propagates high-level semantic information downward, ensuring that all layers contribute to detecting objects of various sizes. This involves: Downsampling through Convolution: The high-resolution P3 feature map is downsampled via a Conv Block (e.g., Block 14) with a stride of 2. This operation reduces spatial resolution while enhancing semantic richness, preparing the features for fusion with deeper layers (e.g., P4 and P5).

Feature Fusion: The downsampled P3 features are concatenated with the P4 features using a Concat Block (Block 20). Similarly, the P4 features are processed and concatenated with the P5 features through Concat Block (Block 21).

C2f Refinement: After each concatenation, the fused feature maps pass through C2f Blocks (e.g., Blocks 18 and 21). These blocks apply convolutional refinements and shortcuts to maintain spatial and semantic balance across scales.

To resume the neck, The PAN-FPN’s bidirectional flow is really important in handling multiscale objects thanks to a Bottom-Up Pathway (That captures spatially detailed information from deeper layers (P5 to P3) through upsampling and concatenation, ensuring smaller objects are accurately detected) and a Top-Down Pathway (That reintroduces high-level semantic information (P3 to P5) through downsampling and concatenation, improving contextual understanding for larger objects.)

Detection Head Overview

The detection head (depicted in the far-right section of Figure 2) is responsible for generating the final predictions, including object classes, bounding boxes, and confidence scores. The detection head leverages the fused feature maps produced by the PAN-FPN to ensure accurate predictions across different object scales. By utilizing multiscale feature maps, the detection head can identify both small and large objects with improved precision. Figure 2 illustrates how the multiscale information flows into the detection head, highlighting the importance of fused feature representations in producing accurate predictions.

In conclusion, the PAN-FPN neck in YOLOv8 plays a pivotal role in enabling efficient multiscale feature fusion, handling intra-class variance, and improving the overall detection capability of the model. Through its bottom-up and top-down pathways the PAN-FPN achieves a balanced integration of spatial and semantic information, which is critical for detecting objects of various sizes. The bidirectional feature fusion mechanism ensures that the network remains robust in complex detection scenarios, making PAN-FPN an indispensable part of YOLOv8’s performance in diverse environments.

D. Our model-Adaptive Path Fusion YOLO (APF-YOLO)

The YOLOv8 model has achieved remarkable advancements in object detection, offering significant improvements in speed and accuracy. However, it faces notable challenges in handling multiscale objects and intra-class variance, particularly under the complex environmental conditions of UAV-based power line inspections. This is primarily because YOLOv8 was not specifically designed to address the unique challenges of such applications. To overcome these limitations, we introduce APF-YOLO (Adaptive Path Fusion YOLO), which focuses on tackling the dual challenges of multiscale object detection and intra-class variance. By incorporating innovative architectural changes and advanced feature fusion strategies, APF-YOLO enhances the detection of both small and large objects while effectively differentiating visually similar objects within the same class. Below, we outline the core methodology and contributions of APF-YOLO:

Multiscale Object Detection

Detecting objects of varying sizes in UAV imagery presents significant challenges due to the wide range of spatial scales. Small objects require precise attention to fine-grained details, while larger objects demand an understanding of broader contextual relationships. To address these challenges, APF-YOLO introduces the Adaptive Feature Alignment Module (AFAM), a key innovation designed to dynamically align and fuse feature maps from three consecutive layers (e.g., P3, P4, P5).

As shown in Figures 3 and 4, AFAM uses a flexible Feature Alignment process that adjusts the three input feature resolutions to a common scale, with the mid-layer (e.g., P4) serving as the reference resolution. This dynamic adjustment ensures that multiscale features are effectively combined, regardless of the specific input layers. The intuition behind Feature Alignment stems from the observation that feature maps from different layers inherently differ in resolution and semantic content. High-resolution maps (e.g., P3) capture fine spatial details, while low-resolution maps (e.g., P5) provide broader semantic context. Combining these directly without alignment can distort the fused representation, as the differing spatial scales introduce inconsistencies. To address this, AConv, an effective downsampling method from YOLOv9 authors, is used to reduce the resolution of higher-resolution inputs (e.g., P3) while preserving critical spatial features. Simultaneously, upsampling techniques are applied to low-resolution maps (e.g., P5) to bring them to the mid-layer’s resolution. This alignment step not only ensures consistency but also enables seamless fusion of features across scales.

Figure 3. Schematic comparison of YOLOv8 Neck and APF-YOLO Neck.

Figure 4. Architecture of the Adaptive Feature Alignment Module (AFAM) with Local and Global Pathway.

Once aligned, the feature maps are processed through two complementary pathways: the local pathway and the global pathway. The local pathway focuses on fine-grained details from the highest-resolution input (e.g., P3). It applies Depth-Wise Convolutions (DW, k=3) to capture localized patterns efficiently and Efficient Local Attention (ELA, k=3) to refine these features by selectively enhancing critical spatial details. The intuition behind the local pathway is that small objects, such as insulators captured from a distant perspective, require precise spatial detail preservation, which this pathway is explicitly designed to provide. In contrast, the global pathway processes the semantically rich but lower-resolution inputs (e.g., P4 and P5) to capture broader contextual relationships. This pathway starts with Depth-Wise Convolutions (DW, k=11) to expand the receptive field, capturing structural and spatial information across the scene. The global pathway then applies Efficient Channel Attention (ECA) to emphasize the most semantically meaningful feature channels, ensuring that high-level structural attributes are prioritized. Finally, Context Anchor Attention (CAA) refines these features by focusing on spatial and contextual relationships, enabling the model to understand how different parts of the object and its surroundings interact. The intuition behind this pathway is that larger objects and their contexts require broad, high-level understanding that complements the localized detail captured in the local pathway.

The outputs from both pathways are fused through concatenation, creating a unified feature representation that balances spatial precision and semantic richness. This fusion ensures that small objects are detected with high-resolution detail while larger objects and their contextual relationships are captured through broader semantic understanding.

By dynamically aligning feature maps and combining complementary pathways, AFAM is designed to handle multiscale detection challenges effectively, making it particularly suitable for UAV-based applications like power line inspections, where detecting objects of varying sizes and understanding their spatial context are paramount.

To further improve small object detection, we introduced a direct pathway from P2 (backbone) to the neck (Refer to Figure 3). By incorporating this high-resolution feature map, the model gains access to additional fine-grained information, enabling it to better identify and differentiate small objects in complex UAV imagery.

Intra-Class Variance Handling

Intra-class variance, where objects within the same class differ subtly in material or texture (e.g., glass, porcelain, and composite insulators), presents a significant challenge in object detection. These variations in material properties often arise from manufacturing differences or environmental exposure, yet the insulators retain a consistent global shape across classes. This shared structural feature serves as a key focus for addressing intra-class variance.

The Adaptive Feature Alignment Module (AFAM), specifically its global pathway, is designed to leverage this insight by emphasizing high-level semantic information extracted from deeper network layers (e.g., Pi and Pi+1). These layers inherently capture critical structural features such as shape and form, which are essential for distinguishing insulators that share similar appearances but vary in finer material-specific details. To preserve and refine this information, the global pathway begins with a Depth-Wise Convolution (DW, k=11), which expands the receptive field to capture broader spatial relationships. This operation ensures that global features representing the object’s overall shape and structural integrity are prioritized in the feature representation. The intuition behind this design is rooted in the idea that global shape consistency provides a robust framework for distinguishing objects within the same class, even when their textures or fine details differ. The global pathway complements this approach by incorporating advanced attention mechanisms, specifically Efficient Channel Attention (ECA) and Context Anchor Attention (CAA). These mechanisms are designed to address the subtle variations within a class that can manifest in specific feature channels or spatial regions. ECA selectively emphasizes the most relevant feature channels, ensuring that the representation focuses on high-level semantic characteristics like shape and structure while filtering out irrelevant or noisy details. This helps strengthen the representation of shared attributes across the class, reducing the risk of misclassification caused by misleading material-specific features. Meanwhile, CAA operates by capturing meaningful spatial and contextual relationships within the feature maps, ensuring that the global pathway effectively understands how different parts of the object relate to each other and their surroundings.

This spatial refinement is particularly valuable for objects like insulators, which often appear in complex or cluttered environments where contextual information is critical. The outputs of the global pathway, refined through DW, ECA, and CAA, focus on the global shape and semantic consistency of the objects. However, addressing intra-class variance with multi scale objects also requires integrating material-specific granularity. For this purpose, the refined outputs from the global pathway are fused with those from the local pathway, which captures fine-grained material details using high-resolution inputs (e.g., P2 or P3, depending on the scenario). The local pathway complements the global pathway by focusing on features that highlight differences in texture or material properties.

By leveraging these complementary pathways, APF-YOLO is designed to address the challenges of intra-class variance in UAV-based power line inspections, where insulators of different materials must be reliably detected. Future evaluations will assess how effectively this design meets its intended goals in real-world applications, focusing on its ability to balance global consistency and localized detail extraction.

More technical details regarding these components and their implementation can be found in the subsequent sections of this paper.

1) Depth-Wise Convolution (DW)

Depth-Wise Convolution (DW),³¹ as illustrated in Figure 5, offers a computationally efficient alternative to traditional convolution by processing each channel independently with separate filters. Unlike standard convolution, where a single kernel operates across all channels simultaneously (left side of the figure), DW applies a distinct K×K kernel to each channel individually and stacks the outputs (right side of the figure). This significantly reduces computational cost while retaining channel-specific features. For an input feature map of size H×W×C, the computational cost of DW is K² ×H×W×C, compared to K²×H×W×C×N for standard convolution, where N is the number of output channels. In APF-YOLO, DW is applied with kernel sizes k=3 and k = 11, tailored for the distinct roles of the local and global pathways within the AFAM. The local pathway, designed to process high-resolution feature maps, uses a k=3 kernel. As shown in the figure, the smaller kernel size focuses on localized patterns while maintaining computational efficiency. This is critical for detecting small objects where fine-grained spatial detail is essential. On the other hand, the global pathway processes low-resolution feature maps using a k=11 kernel to expand the receptive field and capture broader spatial relationships. As depicted in the figure, the larger kernel size enables the model to aggregate information over a wider area, which is crucial for understanding the structure and context of larger objects. This is particularly important for addressing intra-class variance in insulators (e.g., glass, porcelain, and composite) by focusing on the global shape and form. Figure 5 illustrates how DW separates operations across channels (right) compared to the combined processing of all channels in standard convolution (left). The use of k=3 in the local pathway ensures high-resolution features remain precise, while k=11 in the global pathway emphasizes semantic context. This dual-kernel approach reflects the intuition behind APF-YOLO’s design. Together, they balance localized precision and contextual understanding, enabling the model to handle multiscale object detection and intra-class variance effectively.

Figure 5. Schematic comparison of Standard Convolution and Depth-Wise Convolution.

2) AConv

The AConv module¹⁸ is designed specifically for downsampling to facilitate feature alignment in multiscale architectures like APF-YOLO.As shown in the Figure 6, AConv combines max pooling, which reduces the spatial resolution of the input while preserving dominant features, and a 3×2 convolution with stride 2, which processes spatial information during downsampling. Additionally, two 1×1 convolutions are used to refine and adjust channel dimensions before the outputs are concatenated. The purpose of AConv in APF-YOLO is to resize the feature maps from the P(i−1) layer to match the dimensions of the mid-layer P(i), enabling seamless integration in the Feature Alignment stage.

Figure 6. Schematic representation of Aconv.

3) ELA

The Efficient Local Attention (ELA) module³² is designed to refine and selectively emphasize fine-grained spatial features critical for accurately localizing small objects like insulators. It is strategically placed after the Depth-Wise Convolution (DW, k=3) in the local pathway to build upon the initial spatial features extracted by the convolution. The DW convolution, with its $(\mathbf{3}\mathit{\times }\mathbf{3})$ kernel, serves as the first stage, capturing localized spatial patterns and contours in a computationally efficient manner. This operation is particularly effective in extracting detailed features, such as the precise edges and shapes that are vital for localizing small-scale objects. However, DW operates uniformly across all spatial regions and channels, which may result in less focus on the most relevant localized features. To address this, ELA follows DW to enhance the extracted features by incorporating a selective attention mechanism. As shown in Figure 7, ELA begins with X and Y average pooling operations, which aggregate spatial information along the width and height dimensions, respectively, creating condensed intermediate representations that capture essential local spatial dependencies. These representations are then processed through $(\mathbf{1}\mathit{\times }\mathbf{1})$ convolutions, enabling channel-wise interactions that enrich the feature representation while preserving critical spatial details. Group Normalization stabilizes the output, and a sigmoid activation function generates attention weights that are applied to the original feature maps. This process ensures that the most relevant spatial features are amplified while irrelevant or redundant information is suppressed. The intuition behind this design lies in the complementary roles of DW and ELA. While DW captures fine-grained spatial patterns, ELA refines these patterns by selectively focusing on the most important localized features. This sequential approach enables the local pathway in the Adaptive Feature Alignment Module (AFAM) to effectively combine spatial precision from DW with the selective focus of ELA. By doing so, APF-YOLO is designed to address the challenges of small object localization, ensuring that fine-grained spatial details are accurately captured and emphasized. This makes it particularly well-suited for UAV-based power line inspections, where precise localization of insulators across varying scales is a critical challenge.

Figure 7. Diagrammatic representation of Efficient Local Attention (ELA).

4) ECA

The Efficient Channel Attention (ECA) module³³ is designed to optimize attention across feature map channels, enabling the model to focus on the most informative features while suppressing less relevant ones. As shown in the Figure 8, ECA dynamically learns the importance of each channel and applies adaptive weighting to enhance feature representation. This capability is particularly critical for capturing high-level semantic features, essential for handling intra-class variations, such as differentiating between glass, porcelain, and composite insulators. ECA begins with Global Average Pooling (GAP), which aggregates spatial information $(\mathit{H}\mathit{\times }\mathit{W})$ across each channel of the input feature map. This operation produces a condensed representation of size $(\mathit{C}\mathit{\times }\mathbf{1})$ , where $\left(\mathit{C}\right)$ is the number of channels, summarizing the contribution of each channel. Following this, ECA applies a $(\mathbf{1}\mathit{\times }\mathbf{1})$ convolution to learn inter-channel dependencies, with an adaptive kernel size determined $\mathit{k}\mathit{=}\mathit{\psi }\left(\mathit{C}\right)$ , where $\left(\mathit{\psi }\left(\mathit{C}\right)\right)$ is a mapping function. This adaptive mechanism is especially critical in lower layers, such as P5, which contain a large number of feature maps (i.e., channels). In these layers, the abundance of channels increases the likelihood of redundant or irrelevant information. Channel attention provided by ECA ensures that the model focuses on the most relevant features, filtering out less useful information. After the convolution, the output is passed through a sigmoid activation function, generating channel-wise weights that scale the original feature maps. This recalibration is expressed as $(\mathit{Y}\mathit{=}\mathit{X}\otimes \mathit{s})\mathit{,}$ where X represents the input feature map, s is the learned channel weight, and $(\otimes )$ denotes element-wise multiplication. By assigning greater focus to channels carrying significant semantic information and reducing the influence of irrelevant or noisy channels, ECA enhances the overall representation of feature maps. The ECA module follows the Depth-Wise Convolution (DW, k=11) in the global pathway for a specific reason. The DW convolution, with its large kernel size, captures broad spatial relationships critical for preserving the global shape and structure of objects. However, DW processes each channel independently and does not address inter-channel interactions. In deeper layers like $\left(\mathit{P}\mathbf{5}\right)\mathit{,}$ which contain a high number of feature maps, this lack of channel interaction can dilute the representation’s efficiency. ECA complements DW by recalibrating the importance of each channel, ensuring that only the most relevant features are emphasized while redundant channels are suppressed. This sequence of operations enhances the global pathway’s ability to handle intra-class variance by focusing on structural and contextual features that are semantically meaningful.

Figure 8. Diagrammatic representation of Efficient Channel Attention (ECA).

In APF-YOLO ’s Adaptive Feature Alignment Module (AFAM), ECA plays a crucial role in refining channel-level representations, especially in layers with a high channel count. By leveraging ECA after DW, the global pathway is designed to balance spatial relationships captured by DW with the selective focus of channel attention, ensuring that the model can emphasize critical semantic features while reducing noise. This design ensures that APF-YOLO is meant to effectively handle complex scenarios, such as distinguishing between objects with subtle material differences, while preserving global structural integrity.

5) CAA

The Context Anchor Attention (CAA) module³⁴ introduces a context-aware mechanism designed to enhance the model’s ability to understand spatial relationships and interactions within a scene. This capability is particularly critical in scenarios involving insulators, which typically exhibit elongated rectangular shapes. As shown in the Figure 10, the bounding box annotations reveal that insulators predominantly align along horizontal and vertical axes. This characteristic directly informs the design of CAA, which employs depth-wise convolutions with kernels explicitly tailored for vertical $\left(\right(\mathbf{1}\mathit{\times }(\mathbf{11}\mathit{+}\mathbf{2}\mathit{n})\left)\right)$ and horizontal $\left(\right((\mathbf{11}\mathit{+}\mathbf{2}\mathit{n})\mathit{\times }\mathbf{1}\left)\right)\mathit{\text{contexts}}$ . By expanding the receptive field in these directions, CAA enables the model to analyze spatial dependencies and interactions specific to insulators’ orientation and geometry. As shown in the Figure 9, The CAA module begins by applying Average Pooling to aggregate spatial information from the input feature maps, reducing dimensions while retaining critical channel-wise information. The pooled features are processed through a $(\mathbf{1}\mathit{\times }\mathbf{1})$ convolution, generating context anchors that model inter-channel relationships and lay the foundation for contextual refinement.

Figure 9. Diagrammatic representation of Context Anchor Attention (CAA).

Figure 10. Visualization of Dataset Bounding Box Labels.

Following this, the vertical and horizontal depth-wise convolutions are applied to capture directional spatial relationships. These operations ensure that the model can effectively learn patterns corresponding to the insulator’s shape and context, such as alignment with power lines or placement within complex backgrounds. The outputs from the convolutions are refined through a final $(\mathbf{1}\mathit{\times }\mathbf{1})$ convolution and normalized with a sigmoid activation function, producing attention weights that prioritize the most contextually relevant features. These weights enhance the model’s focus on spatial relationships critical for distinguishing insulators even when partially obscured or fragmented. The ability to analyze the insulator’s context along its natural horizontal and vertical axes ensures that spatial integrity is maintained, making CAA particularly effective in scenarios where accurate localization and classification depend on understanding the object’s surroundings. CAA complements the preceding Efficient Channel Attention (ECA) module in the global pathway. While ECA recalibrates channel relevance to prioritize semantically meaningful features, it does not address spatial dependencies. CAA builds on ECA by introducing spatial context awareness, refining the model’s ability to understand an insulator’s position and alignment within its scene. Together, these modules create a sequential process where features are first optimized for channel relevance and then enriched with spatial and contextual refinements. As depicted in the Figure 10, the rectangular shape and alignment of insulators strongly support the rationale for CAA’s design. Its focus on vertical and horizontal spatial relationships ensures that the model captures critical contextual information necessary for accurate detection, even when insulators are partially occluded or split due to background interference. This approach makes CAA well-suited for UAV-based power line inspections, where understanding both object structure and its spatial context is essential for precise localization and classification. In summary, APF-YOLO is built directly on the YOLOv8 architecture, with modifications made solely to the neck component to enhance multiscale feature fusion and address intra-class variance. The backbone, detection head, loss functions, and other components remain unchanged, leveraging YOLOv8’s inherent strengths while focusing improvements where they are most needed.

E. Dataset description

For our study, The Merged Public Insulator Dataset (MPID)⁴⁷ was constructed to address the need for a diverse, high-quality dataset specifically tailored for UAV-based high-voltage power line inspections. This dataset reflects real-world variability and provides a robust foundation for developing advanced object detection models, particularly those addressing multiscale object detection and intra-class variance challenges. Comprising 4807 images and 7850 labeled instances of three insulator types—glass, porcelain, and composite—as detailed in Table 3, MPID integrates multiple publicly available datasets from different regions worldwide as detailed in Table 2. This ensures a rich diversity of environmental and inspection scenarios, making it a practical and generalizable resource for power line inspection research. MPID includes images captured under diverse conditions, such as varying times of day, weather environments, lighting conditions, and camera angles, replicating the challenges faced during UAV inspections. The dataset also provides images from different distances and perspectives: (a) distant views, (b) full views, and (c) close-ups, as illustrated in Figure 11. This variability in viewpoints ensures that MPID reflects real-world conditions, addressing the inherent challenges posed by insulator scale variability in UAV imagery. Furthermore, the width-height distribution of insulators, visualized in Figure 12, highlights the multiscale problem inherent in MPID. The range of bounding box sizes indicates a broad spectrum of object dimensions, ensuring that models trained on MPID must detect both small and large insulators effectively (The darker blue regions indicate a higher concentration of insulators with those particular width-height ratios). This is particularly relevant for APF-YOLO, which aims to address multiscale detection by preserving critical information. MPID emphasizes three distinct types of insulators—glass, porcelain, and composite—which share a consistent global shape but differ in material and texture, introducing significant intra-class variance. Glass insulators are characterized by their transparency, porcelain insulators exhibit a ceramic texture, and composite insulators feature polymeric materials. This diversity challenges detection models to preserve semantic information while distinguishing between these visually similar objects. The inclusion of this variety makes MPID particularly relevant for developing models capable of managing intra-class variability. The dataset construction involved a semi-automatic annotation process to ensure high-quality labels. After filtering irrelevant images, pre-trained object detection models were used to generate initial bounding boxes, which were then manually refined using the Roboflow platform.⁴⁴ Annotators ensured that bounding boxes tightly encapsulated each insulator and accurately identified its type. The annotations were formatted in the YOLOv8 annotation standard, enabling seamless compatibility with YOLO-based models. To standardize the dataset, all images were resized to 640 × 640 pixels, the default input size for YOLO architectures. Unlike traditional dataset preparation processes, no manual data augmentation was applied during MPID construction. Instead, YOLOv8’s built-in augmentation capabilities—such as random scaling, rotation, and color adjustments—were leveraged during the training pipeline. This ensures that the model benefits from diverse variations in data without requiring explicit augmentation during dataset preparation. The use of YOLOv8’s augmentation mechanisms reduces preprocessing efforts while enhancing generalization during training. Furthermore, a typical 80-20 train-validation split was applied to the dataset to facilitate model evaluation. By offering a diverse and realistic dataset, MPID supports the development of advanced detection models. Its focus on multiscale variability and intra-class diversity ensures that models trained on it are equipped to handle the complexities of real-world UAV-based power line inspections.

Figure 11. Distribution of Object Bounding Box Aspect Ratios (Height vs. Width).

Figure 12. Sample images of the MPID.

We propose 3 types of insulators, (1) glass (2) composite and (3) porcelain. With different distances of shooting by the drone (a) distant (b) full view (c) close up.

A. Environment configuration

The training and validation processes for this study were conducted in a carefully designed computing environment tailored to the computational demands of real-time object detection tasks. The experiments utilized a virtual machine equipped with a Quadro P6000 GPU, offering 24 GB of VRAM. This GPU’s high memory capacity and parallel processing capabilities ensured efficient handling of the large multiscale dataset and computationally intensive training of deep learning models. CUDA version 12.5 was employed to leverage GPU acceleration, streamlining algebraic operations critical for optimizing neural network training.

The software environment was built on Windows 10, chosen for its stability and compatibility with diverse machine learning tools. Python 3.9.10 served as the programming language, enabling seamless integration with the PyTorch version 2.2.2+cu121 which was selected for its flexibility and robust support for neural network development, making it particularly suitable for adapting and fine-tuning YOLO models. Model development and training were further facilitated by Ultralytics’ tools, optimized for YOLOv8 models and designed to enhance both speed and accuracy in detection tasks.

To ensure efficient management and reproducibility of experiments, the MLOps platform ClearML⁴⁵ was employed. ClearML provided a centralized system for tracking experiments, monitoring metrics, and collaborating across development stages, supporting the study’s iterative experimentation process.

This computing environment not only ensured smooth training and validation but also provided a foundation for reproducibility, scalability, and efficiency throughout the experimental process.

B. Models

This section outlines the models chosen for comparison, focusing specifically on state-of-the-art (SOTA) neck configurations within the YOLO family. These configurations are evaluated for their ability to address key challenges in object detection, such as multiscale detection and intra-class variance. By comparing a variety of neck designs, the study aims to understand their strengths and weaknesses in handling these complexities.

A diverse set of models is included in the analysis, each representing unique design philosophies for improving feature fusion and context understanding. These models include YOLOv8n_baseline (default YOLOv8 neck), YOLOv8n-p2 (incorporating the P2 layer for small object detection), YOLOv8n-MAFPN, YOLOv8n-HSFPN, YOLOv8n-goldyolo, YOLOv8n-CFPT-P2345, YOLOv8n-bifpn, and YOLOv8n-ASF-P2. For additional perspective, the study also considers YOLOv6n, YOLOv5n, YOLOv9-t, and YOLOv10n models, representing variations in architecture and size.

The models chosen are primarily in the “n” (nano) size category, which are lightweight configurations optimized for edge devices and consumer GPUs. This focus ensures that the analysis remains relevant to real-world applications, such as UAV-based power line inspections, where computational resources are often limited. Larger models, while offering higher precision due to increased parameter counts, are typically excluded as they do not align with the operational constraints of real-time UAV integration. The emphasis on nano models also reflects their suitability for deployment on devices with restricted hardware capacity, balancing efficiency and accuracy.

C. Metrics

To provide a comprehensive assessment of model performance, it is generally considered appropriate to utilize several different evaluation metrics. For the evaluation of insulator defect detection models, we employ a suite of key metrics to ensure a thorough and nuanced understanding of performance. These metrics include Precision, Recall, Mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds, inference time, and Giga Floating Point Operations per Second (GFLOPS). Each metric provides distinct insights into the model’s effectiveness and efficiency.

Precision P measures the accuracy of the model’s positive predictions and is defined as:

Recall R assesses the model’s capability to detect all relevant defect instances. It is computed as:

Average Precision (AP) evaluates the model’s performance across different recall levels by integrating the precision-recall curve. The AP is computed as:

Mean Average Precision (mAP) extends AP to multiple classes by averaging AP values across all categories. It is defined as:

In addition to mAP, we assess the model using Average Precision (AP) at specific Intersection over Union (IoU) thresholds. We calculate AP at IoU = 0.5 (denoted as AP@0.5) and average AP across IoU thresholds from 0.5 to 0.95 in increments of 0.05 (denoted as AP@0.5:0.95). These metrics provide a detailed evaluation of the model’s performance with varying degrees of overlap between predicted and ground truth bounding boxes.

Inference time measures the duration required for the model to process an image and produce predictions. This metric is critical for real-time applications where rapid processing is essential. Inference time is expressed in milliseconds and reflects the model’s efficiency in practical scenarios.

GFLOPS quantifies the computational complexity of the model by indicating the number of floating-point operations performed per second. This metric helps assess the model’s computational demands and is crucial for understanding its performance capabilities. Higher GFLOPS values suggest more complex models, which may impact real-time processing.

These metrics collectively offer a thorough evaluation framework for assessing the performance, efficiency, and practicality of object defect detection models.

D. Training process

Models are trained using standard parameters and hyperparameters recommended by the original authors, accessible via the Ultralytics platform.³⁰ This approach ensures consistency with established research standards, providing a fair and unbiased basis for comparison. To ensure unbiased performance evaluation, models are trained from scratch. For this study, we focus on various neck configurations, including YOLOv9-t, YOLOv8n_baseline, YOLOv8n-p2, YOLOv8n-MAFPN, YOLOv8n-HSFPN, YOLOv8n-goldyolo, YOLOv8n-CFPT-P2345, YOLOv8n-bifpn, YOLOv8n-ASF-P2, YOLOv6n, YOLOv5n, and YOLOv10n.

The training process involves 50 epochs with a batch size of 32, and the input image size is set to 640 pixels. Data augmentation techniques are automatically applied during training, enhancing the model’s robustness and enabling it to generalize better across varying conditions. The use of data augmentation is crucial, as it simulates a diverse set of scenarios, helping to improve the model’s ability to handle multiscale detection and intra-class variance. The total training duration for all models is approximately 11 hours, allowing for thorough evaluation over the specified epochs. Each model is monitored closely throughout the training process to track the progression of model loss functions. Initial observations indicate that models do not exhibit signs of overfitting, and the loss functions (DFL loss and box regression loss) reliably converge to low values, reflecting effective training practices.

Given that the dataset contains a single class (insulators), classification loss (Cls Loss) is not monitored, as it is not relevant in this context.

To visualize the training efficacy, performance curves are generated, demonstrating stability and convergence around the 50-epoch step. While we primarily focus on the mean Average Precision (mAP) metric to minimize clutter in results presentation, other evaluation metrics exhibit similar trends, reinforcing the good training phase. These findings are illustrated in the accompanying Figure 13, providing a clear picture of the training outcomes.

Figure 13. Training Stability and Validation Convergence Across YOLO Variants: Assessing Overfitting and Epoch Efficiency.

All hyperparameters used in this study can be found in the open-source code provided through this link: [https://github.com/phd-benel/APF-YOLOv8].

1) Comparative analysis of our contribution against recent advancements

The current evaluation aims to analyze the performance of state-of-the-art (SOTA) models for insulator detection, a complex task involving multi-scale detection, small object sizes, and intra-class invariance. These challenges make it difficult for models to achieve exceptionally high performance (e.g., near 99% accuracy) across all metrics, as seen in the Table 4. The analysis covers key metrics such as mAP 50, mAP 50-95, precision, recall, and computational efficiency, shedding light on how each model balances detection performance with computational cost.

Overview of Model Performance

The results indicate that APF-YOLO achieves the best overall performance, with a mAP 50 of 0.9241, mAP 50-95 of 0.6801, precision of 0.9243, and recall of 0.8561. These results highlight its effectiveness in detecting insulators with diverse scales and intra-class variations. However, these gains come with increased computational requirements, as indicated by 12.779 GFLOPs, 3.6 million parameters, and an inference speed of 5.163 ms. Despite the additional computational demands, APF-YOLO demonstrates an optimal balance between accuracy and resource consumption, making it suitable for multi-scale detection and complex object variations.

Analysis with Specific Metrics

Compared to yolov8n_baseline, which serves as a standard benchmark, APF-YOLO achieves a 5.5% increase in mAP 50 (0.9241 vs. 0.8760) and a 6.9% increase in mAP 50-95 (0.6801 vs. 0.6365). It also achieves a 3.4% improvement in precision (0.9243 vs. 0.8938) and a 5.5% improvement in recall (0.8561 vs. 0.8111). These gains suggest that APF-YOLO effectively addresses the complexities of detecting insulators across multiple scales and types. However, this comes at the cost of 56% higher GFLOPs, 18.7% more parameters, and a 92% slower inference speed compared to yolov8n_baseline. This trade-off indicates that while accuracy significantly improves, computational efficiency decreases, highlighting a challenge for real-time applications.

Comparative Analysis Against Competitors

When compared to yolov8n-p2, APF-YOLO achieves a 0.74% higher mAP 50 (0.9241 vs. 0.9173) and a 2.72% increase in mAP 50-95 (0.6801 vs. 0.6621). While precision is slightly lower (0.9243 vs. 0.9273), APF-YOLO achieves 1.23% higher recall (0.8561 vs. 0.8456), suggesting better overall coverage of detected objects. The computational cost is also slightly higher, with 3.4% more GFLOPs, 22% more parameters, and 33% slower inference speed. Despite this, APF-YOLO achieves better generalization and recall, making it more suitable for comprehensive detection tasks.

Other competitors like Yolov10n, Yolov8n-MAFPN, Yolov8n-HSFPN, and Yolov8n-goldyolo show varied performance levels. For instance, Yolov10n achieves a 6.2% lower mAP 50 (0.9241 vs. 0.8700) and a 9.1% lower mAP 50-95 (0.6801 vs. 0.6182). Similarly, Yolov8n-HSFPN has 5.8% lower mAP 50 (0.8731) and 7.9% lower mAP 50-95 (0.6267), along with a 3.6% lower precision. Meanwhile, Yolov8n-MAFPN and Yolov8n-goldyolo also record lower performance in both mAP metrics and precision, with Yolov8n-MAFPN lagging behind by 3.9% in mAP 50 and 4.8% in mAP 50-95, and Yolov8n-goldyolo showing 6.2% lower mAP 50 and 10.3% lower mAP 50-95. These results underscore Our model’s competitive advantage in terms of accuracy, albeit with a higher computational cost.

The results depicted in Figure 14 illustrate various challenging scenarios and highlight instances where APF-YOLO demonstrates strong performance compared to other models. In the first two rows, APF-YOLO shows precise bounding box placement for insulators of different types (glass in the first row and composite in the second), maintaining tight and accurate localization. Competing models, such as YOLOv8-baseline and YOLOv8-ASF-P2, either fail to detect these insulators entirely or produce overly large bounding boxes, which could limit their practical effectiveness.

Figure 14. Illustrative Comparison of Model Predictions on the Validation Set.

The results are displayed with bounding boxes, where the colors are arbitrary and do not convey any specific meaning. The ground truth represents the actual location of the object, serving as a reference for evaluation.

In the third row, showing insulators captured from a distant perspective, APF-YOLO detects four insulators compared to three detected by YOLOv8-p2, indicating its potential for superior handling of distant objects. While no model detects all insulators in this scenario, APF-YOLO shows an advantage in managing this challenging case.

The fourth and fifth rows highlight the model’s ability to detect partially occluded insulators behind metallic pylons. APF-YOLO accurately identifies these occluded objects in their entirety, a critical feature for real-world deployments. Competitors like YOLOv8-p2 can’t manage this occlusion detection, further demonstrating the limitations of these alternative approaches.

In the sixth and seventh rows, APF-YOLO stands out in particularly extreme scenarios. For example, in row six, it successfully detects a heavily occluded insulator in the top-left corner, a scenario where other models struggle. Similarly, in row seven, APF-YOLO identifies an insulator (in the top) within a complex background, a task other models fail to achieve. These cases underscore its ability to handle extreme conditions, which are critical for UAV-based power line inspections.

Overall, Figure 14 illustrates that APF-YOLO performs effectively in a range of scenarios, including multiscale detection, occlusions, and challenging backgrounds. While the results do not suggest that the model is universally superior in all cases, they demonstrate its capability to handle extreme conditions—a key requirement for practical deployment in real-world inspection tasks.

Explanation of Robustness or Weaknesses in Performance

APF-YOLO achieves superior results across key metrics such as mAP 50 (0.9241), mAP 50-95 (0.6801), precision (0.9243), and recall (0.8561), reflecting its robustness in addressing multi-scale detection, small object recognition, and intra-class variance among different insulator types, including glass, porcelain, and composite. These strong results can be attributed to the innovative design of the AFAM module, which integrates specialized paths for handling both global semantic information and small object features.

In the AFAM module, one of the key design elements is the pathway dedicated to global semantic contextual information. This pathway is crucial for managing intra-class variance, as the global shape of insulators tends to be similar (e.g., disk-like shapes), despite variations in material type or structure. The module starts with a depthwise convolution (DW) with a kernel size of 11, which captures extensive contextual information across larger receptive fields. This operation is particularly effective for extracting global semantic features that help the model generalize across different insulator types. The large kernel size aligns well with detecting larger object sizes, which are generally found in deeper network layers.

Following the DW-11, the module incorporates Efficient Channel Attention (ECA) to refine channel-wise features. The deeper layers of the network contain a greater number of channels, making ECA especially effective in preserving relevant features and suppressing noise. By applying ECA after global semantic extraction, the model optimizes feature representation, ensuring that the most significant channels contribute to the detection of both large objects and intra-class variations.

The next component in the global pathway is the Context Anchor Attention (CAA). By leveraging vertical and horizontal convolutions, CAA enhances the model’s understanding of the spatial context. This mechanism further improves the model’s ability to differentiate between similar shapes by enriching global context representation, which is critical for accurately detecting insulators with varying sizes and orientations. The combination of these operations in the global pathway supports high precision and recall, demonstrating the model’s capability to capture complex spatial relationships and maintain generalization across different IoU thresholds. In parallel, the AFAM module also includes a pathway specifically designed for small object detection, addressing the challenges of multi-scale detection. This path starts with depthwise convolution (DW) with a kernel size of 3, which targets smaller receptive fields and is ideal for extracting fine-grained features. The use of a smaller kernel size is critical for detecting small objects, which often require precise localization. To enhance the detection of these smaller features, Efficient Local Attention (ELA) is applied, focusing on local spatial information while maintaining channel richness. This ensures accurate feature localization without compromising overall detection performance, contributing to Our model’s high mAP 50-95 and recall rates.

The architectural choices in APF-YOLO yield high detection performance, but they also introduce certain computational costs. The three-input processing of the AFAM module, along with the integration of large kernels, multiple attention mechanisms, and separate paths for global and local features, increases the computational load. This design contributes to the higher GFLOPs (12.779), a greater number of parameters (3.6 million), and a slower inference speed (5.163 ms), compared to conventional models that typically process a single input feature. The complexity of handling both global and local contexts simultaneously, while advantageous for accuracy, results in longer processing times and greater resource consumption.

In summary, APF-YOLO achieves the highest accuracy among the evaluated SOTA models, demonstrating superior performance in terms of mAP 50, mAP 50-95, and recall. It effectively balances precision and recall, making it a strong choice for insulator detection tasks that involve small objects and intra-class variations by incorporating specialized attention mechanisms and parallel paths for global and local feature extraction.

Future research could focus on optimizing APF-YOLO inference speed and reducing resource consumption. Strategies such as model pruning, quantization, knowledge distillation, and the use of efficient attention mechanisms could improve its applicability for real-time deployment in resource-constrained environments, while maintaining high accuracy across all metrics. Additionally, the model’s architecture makes it a promising candidate for scaling applications, such as cloud computing, where large-scale models can leverage increased computational capacity without compromising detection performance. While the computational costs are higher due to the complex architecture, the trade-off is justified by the model’s strong performance across all key metrics.

2) Ablation study

The primary goal of this ablation study is to evaluate how well our proposed model scales from nano to small configurations compared to baseline models, YOLOv8 and YOLOv8-P2. Scaling is a critical feature for models used in edge computing and UAV-based detection, as it enables adapting model complexity to different deployment scenarios while maintaining robust performance. By isolating the nano and small configurations in this study, we assess the effectiveness of APF-YOLO in leveraging additional computational resources and parameters to improve detection performance relative to baseline models.

The scaling performance of APF-YOLO is evident in the substantial gains observed in both mAP 50 and mAP 50-95 metrics when transitioning from nano to small configurations. For instance, APF-YOLO nano achieves a mAP 50 of 0.9241, which improves to 0.9456 in APF-YOLO small, a relative increase of 2.3%. This scaling gain is more pronounced than that observed in YOLOv8-P2, where mAP 50 increases from 0.9173 in the nano version to 0.9330 in the small version, reflecting a smaller relative improvement of 1.7%. Similarly, YOLOv8-baseline shows a smaller scaling improvement, with mAP 50 increasing from 0.8760 in the nano version to 0.8937 in the small version.

The mAP 50-95 metric, which evaluates detection robustness across varying IoU thresholds, follows a similar trend. APF-YOLO achieves a gain of 0.0375 (from 0.6801 to 0.7176) when scaled from nano to small, significantly outperforming YOLOv8-P2 (0.0319 gain) and YOLOv8-baseline (0.0261 gain). These results indicate that APF-YOLO effectively utilizes additional parameters and computational resources to generalize better across scales, demonstrating superior scalability in performance compared to baseline models.

Precision and recall metrics further highlight the advantages of APF-YOLO in scaling. In the nano configuration, APF-YOLO nano achieves a precision of 0.9243 and a recall of 0.8561. When scaled to APF-YOLO small, these metrics improve to 0.9469 (precision) and 0.8915 (recall), reflecting increases of 2.4% and 3.9%, respectively. These improvements are significantly greater than those observed in YOLOv8-P2, where precision increases by 0.3% and recall by 3.2%, and YOLOv8-baseline, which shows improvements of 2.4% in precision and 1.9% in recall. This suggests that APF-YOLO better leverages scaling to enhance detection reliability, particularly in scenarios with high intra-class variance or multiscale object detection challenges.

While scaling improves performance, it also incurs additional computational costs, as reflected in the GFLOPs and inference speed metrics. For APF-YOLO, GFLOPs increase from 12.779 in the nano configuration to 37.545 in the small configuration, a relative increase of 193%. Similarly, inference speed decreases from 5.163 ms to 9.652 ms, reflecting a 87% slowdown. These changes are comparable to the scaling costs in YOLOv8-P2, where GFLOPs increase from 12.358 to 36.947 (199%) and inference time increases from 3.866 ms to 6.932 ms (79%). YOLOv8-baseline shows similar trends, with GFLOPs increasing by 250% and inference speed decreasing by 69%.

Despite these computational increases, APF-YOLO small consistently outperforms baseline small models in terms of accuracy and robustness, indicating a more efficient utilization of additional resources. This efficiency is crucial for edge computing scenarios, where computational overhead must be justified by substantial gains in detection performance.

The ablation study clearly demonstrates that APF-YOLO scales more effectively than both YOLOv8 and YOLOv8-P2. While all models show performance improvements when scaled, the relative gains in mAP 50, mAP 50-95, precision, and recall are consistently higher in our model. Additionally, the computational cost and inference speed penalties incurred by scaling are comparable across all models, but the performance advantages of APF-YOLO small justify these trade-offs more effectively.

For instance, the mAP 50-95 gain of 0.0375 in APF-YOLO is significantly higher than the gains of 0.0319 and 0.0261 observed in YOLOv8-P2 and YOLOv8-baseline, respectively, even though all models experience similar increases in GFLOPs and decreases in speed. This highlights that APF-YOLO achieves a better balance between performance and computational efficiency, making it a more scalable and practical choice for UAV-based detection applications.

The results of this first ablation study underscore the superior scalability of APF-YOLO compared to YOLOv8 and YOLOv8-P2. By effectively leveraging additional computational resources and parameters, APF-YOLO small achieves the highest performance gains among all tested models, particularly in challenging detection scenarios involving multiscale objects and high intra-class variance. These findings validate the design choices in our architecture, demonstrating its robustness and adaptability for edge computing applications in UAV-based systems. Consequently, APF-YOLO sets a new benchmark for scalable, high-performance object detection in resource-constrained environments.

II. Ablation Study: Optimal Placement of the AFAM Module in the Neck ( Table 6)

The AFAM module is an architectural component aimed at enhancing feature refinement in object detection models. Its placement within the neck plays a crucial role in determining the balance between detection performance and computational efficiency. In this study, we evaluate three different placement strategies for the AFAM module—at the last stage of the neck, at the beginning of the neck, and distributed across both positions (two at the beginning and two at the last). The goal is to identify the optimal placement that maximizes detection performance while maintaining manageable computational costs and inference speed.

Performance Metrics Across Different Configurations

Computational Cost and Efficiency Analysis

The placement of the AFAM module significantly affects the model’s computational efficiency. Placing the module at the beginning minimizes computational costs (12.315 GFLOPs) and keeps inference speed low (5.171 ms), making it the most resource-efficient configuration. However, the performance trade-offs in terms of mAP 50 and precision may not be suitable for applications requiring high detection accuracy.

In contrast, placing the module at the last stage slightly increases computational cost (12.779 GFLOPs) while maintaining a similar inference speed (5.163 ms). This configuration strikes a balance between accuracy and efficiency, making it a practical choice for real-time applications where moderate computational overhead is acceptable.

The distributed placement introduces the highest computational cost (14.071 GFLOPs) and slows inference speed to 6.509 ms. While the additional overhead limits its suitability for resource-constrained environments, the substantial performance improvements in detection metrics make it an attractive choice for accuracy-critical tasks.

The ablation experiments reveal that the distributed placement of the AFAM module—two at the beginning and two at the last stage of the neck—yields the best overall performance. This configuration maximizes detection accuracy (mAP 50: 0.9262, mAP 50-95: 0.6876) and robustness while maintaining competitive precision and recall. Although it incurs higher computational costs and slower inference speeds, the performance gains justify its application in scenarios where accuracy is paramount. For applications prioritizing computational efficiency and real-time processing, the AFAM module’s placement at the last stage provides a viable alternative, balancing accuracy and resource demands. These insights demonstrate the critical impact of architectural decisions on model performance and their implications for edge computing deployments.

III. Ablation Study: Evaluating the Impact of Module Placement on the AFAM Module ( Table 7)

This ablation study investigates the effect of various module placements and configurations within the AFAM module of the APF-YOLO model. Specifically, we explore the role of depthwise convolutions (DW), attention mechanisms, and the inclusion of local and global pathways in enhancing the model’s detection performance. By systematically altering these components, we aim to determine their contribution to the AFAM module’s overall effectiveness.

Comparative Analysis of Configurations

The results indicate that the AFAM module benefits most from configurations that integrate both depthwise convolutions and attention mechanisms. Specifically, the combination of depthwise convolutions and attention mechanisms achieves the best balance between performance metrics and computational efficiency. The following observations stand out:

Best Performance: The configuration with depthwise convolutions and no attention achieves the highest recall (0.8691), making it suitable for applications requiring high coverage of true positives. However, it sacrifices precision, which could lead to more false positives.

Balanced Trade-Off: The configuration with both depthwise convolutions and attention mechanisms achieves a strong balance between precision (0.9194) and recall (0.8526). It demonstrates the effectiveness of combining local feature extraction (via depthwise convolutions) and feature weighting (via attention mechanisms).

Small and Large Path Contributions: Excluding either the small local path or the large context path reduces performance, indicating their complementary roles in handling small and large objects, respectively. The small local path is critical for capturing fine-grained details, while the large context path enhances the model’s ability to interpret global spatial relationships.

The ablation study underscores the importance of combining depthwise convolutions and attention mechanisms within the AFAM module. The inclusion of both small and large pathways is essential for achieving robust multiscale detection. While configurations with depthwise convolutions alone are computationally efficient, the integration of attention mechanisms further refines feature representations, leading to improved detection accuracy and robustness.

For optimal performance in real-world UAV-based detection applications, the full AFAM module—integrating both pathways and leveraging advanced mechanisms—emerges as the most effective design choice.