Receptive-Field-Aware Adaptive Fusion for Multiscale Insulator Defect Detection in Transmission-Line Images

Abstract

Transmission-line inspection requires a single detector to localize complete insulator strings together with much smaller broken-shell and flashover-damaged regions. These targets differ in spatial extent and in their dependence on local detail and surrounding context, which complicates cross-scale feature fusion when accurate bounding boxes are required. This study develops receptive-field-aware path aggregation (PRA) and bidirectional receptive-field-aware aggregation (BRA), combining path aggregation with receptive-field expansion and adaptive scale weighting. Compared with the path aggregation network (PANet), PRA increases mean average precision (mAP) at an intersection-over-union (IoU) threshold of 0.5 (mAP@0.5) by 2.2 percentage points and mAP averaged over IoU thresholds from 0.5 to 0.95 (mAP@0.5: 0.95) by 4.6 points; BRA produces gains of 1.5 and 4.0 points, respectively. The highest mAP@0.5:0.95 of 91.3% is jointly achieved by Swin-T-PRA + Alpha-CIoU and Swin-T-BRA + Alpha-CIoU, with corresponding mAP@0.5 values of 98.6% and 98.5% and model-only inference speeds of 48.5 and 48.3 FPS on an RTX 3080 Ti, respectively. The larger gains across stricter IoU thresholds indicate that PRA/BRA provide greater benefits when more stringent box-overlap criteria are imposed, although the present evaluation does not independently isolate the bounding-box regression mechanism.

1. Introduction

Automatic screening of transmission-line inspection images requires reliable detection of complete insulator strings and localized defects. Insulators provide electrical insulation and mechanical support, while their surface geometry increases the creepage distance and reduces the risk of discharge between conductors at different potentials [1,2]. During long-term service, surface contamination and moisture can form conductive films, reduce the flashover voltage, and promote localized discharge or flashover damage [3,4]. Fractures, cracks, and other visually identifiable broken defects may also appear in transmission-line inspection images and require accurate localization [5]. As illustrated in Figure 1, the resulting inspection task includes three visually distinct classes: insulator strings, broken shells, and flashover-damaged shells. Complete strings are elongated and occupy relatively large image regions, whereas the two defect classes are more localized and rely more heavily on fine appearance cues. A detector must therefore balance broad contextual representation with the preservation of local detail and accurate box localization.

Insulator condition can be assessed using image-based inspection and unmanned aerial vehicle (UAV)-assisted visual analysis [6], together with ultrasonic testing for internal defects [7]. For extensive transmission networks, however, manual inspection is labor-intensive, dependent on operator experience, and difficult to perform consistently in hazardous or inaccessible locations [8,9]. UAVs equipped with imaging sensors are therefore widely used to collect transmission-line images [10]. These inspections can generate large image collections that remain costly to examine manually, particularly when defects occupy only a small image region or have weak visual contrast. Earlier vision-based approaches generally relied on hand-designed image features and rule-based segmentation pipelines. Their dependence on manually specified visual cues limits their ability to accommodate the variation in target scale, viewpoint, illumination, and background found in transmission-line imagery.

Convolutional neural network (CNN)-based detectors have substantially advanced the automatic localization and recognition of insulators and associated defects in aerial inspection images [11], and recent studies continue to explore improved detection architectures, feature representations, and optimization strategies for transmission-line inspection [12]. Ling, et al. [13] combined Faster region-based convolutional neural network (Faster R-CNN) and U-Net to locate self-blast glass insulators in low-signal-to-noise aerial images, while Zhao, et al. [14] incorporated a feature pyramid into Faster R-CNN for insulator recognition under complex backgrounds. Alternative target representations have also been investigated. Wei, et al. [15] represented insulator strings as assembled keypoint chains, and Liu, et al. [16] used parallel branches to locate fault regions and estimate insulator endpoints. Souza, et al. [17] combined a You Only Look Once (YOLO) detector with a ResNet-18 classifier for image-based transmission-line inspection, whereas Wei, et al. [18] examined lightweight edge detection together with cloud-based recognition models. More recent work has extended this line of research toward modern YOLO and real-time transformer architectures. For example, MAP-YOLOv8 integrates targeted architectural modifications for transmission-line insulator defect detection, while improved RT-DETR-based approaches combine attention enhancement, small-object detection layers, and refined boundary-aware feature aggregation [19,20]. Related studies have also examined edge-assisted recognition, cloud–edge collaborative inspection, remote UAV-based component detection, and lightweight power-line defect detection workflows [13,18,21,22,23], although their system-level objectives differ from the image-level detection problem considered here.

Despite this progress, insulator detection in aerial images remains sensitive to heterogeneous backgrounds, viewpoint and illumination changes, target overlap, and limited defect-pixel coverage [17,24,25,26]. Limited annotated data can further constrain data-driven detectors [26,27], and small defect regions tend to lose discriminative detail during feature downsampling [28,29,30]. Outdoor acquisition may additionally introduce noise, blur, fog-like degradation, illumination variation, and contrast changes. Recent UAV-based studies further indicate that sandy and dusty conditions can introduce substantial color bias and visibility degradation, making contextual feature extraction and small-defect localization more difficult [31,32]. Much of the existing work also considers a single defect type, although joint detection of complete strings, broken shells, and flashover-damaged shells requires one model to represent targets with markedly different spatial extents and visual cues. Related studies have adopted multiscale pyramids, receptive-field modules, attention mechanisms, and lightweight detectors, but comparisons often emphasize detection at an intersection-over-union (IoU) threshold of 0.5. This metric alone does not show whether predicted boxes remain accurate at stricter IoU thresholds, and the combined roles of backbone representation, cross-scale fusion, and box-regression loss have not been consistently examined from this perspective. Recent studies [32,33,34] have increasingly adopted modern YOLO variants, transformer-based feature representations, lightweight architectures, and multiscale feature-fusion strategies for insulator defect detection. However, most existing work primarily emphasizes aggregate detection performance, whereas the joint effects of receptive-field expansion, cross-scale feature interaction, and strict IoU localization have received comparatively limited systematic investigation.

This study addresses the problem through receptive-field-aware path aggregation (PRA) and bidirectional receptive-field-aware aggregation (BRA), two fusion necks designed for one-stage insulator detection. Both structures coordinate multi-path feature propagation, multi-branch receptive-field expansion, and adaptive scale weighting so that localized defect cues can be retained while broader context is incorporated for complete insulator strings. Convolutional and vision-transformer backbones provide alternative feature representations, and IoU-based losses are examined as complementary choices for box regression. Cycle-consistent generative adversarial network (CycleGAN)-derived and synthetically fogged images extend the range of appearances used during training; they are not treated as a separate robustness experiment.

The principal contribution lies in the design and evaluation of PRA and BRA for a detection problem in which the target classes have substantially different spatial scales. The two necks integrate established path-aggregation, receptive-field, and adaptive-fusion operations into a task-oriented structure, and their effects are evaluated using mean average precision (mAP) at IoU thresholds of 0.5 and 0.5:0.95. Relative to the corresponding path-aggregation references, the larger improvements at the stricter metric indicate that PRA and BRA provide greater benefits when more stringent box-overlap criteria are imposed. These results characterize high-IoU detection performance rather than independently isolating the bounding-box regression mechanism. Evaluating the necks together with alternative backbones and localization losses further clarifies how feature representation, cross-scale interaction, and boundary regression contribute within the same experimental framework.

Section 2 describes the dataset, training-image preparation, detection framework, PRA/BRA necks, comparative configurations, and evaluation protocol, after which Section 3 presents the experimental results. Section 4 discusses the findings on strict localization, application scope, and limitations, and Section 5 summarizes the main conclusions.

2. Materials and Methods

2.1. Reference Detector and Training Objective

YOLO V5 [35], whose detailed architecture is provided in Figure A1, is used as the reference detector. The cross-stage partial Darknet53 (CSPDarknet53) backbone extracts hierarchical features, and the path aggregation network (PANet) combines them through top-down and bottom-up paths. Prediction heads at P3 (80 × 80), P4 (40 × 40), and P5 (20 × 20) estimate class probabilities, bounding-box coordinates, and object confidence. All subsequent comparisons retain this detection framework while varying the neck, backbone, or localization loss.

The training objective is evaluated at the P3 (80 × 80), P4 (40 × 40), and P5 (20 × 20) prediction scales and contains classification, bounding-box, and confidence terms. For each anchor, binary cross-entropy compares the predicted class probabilities with the ground-truth class labels. The classification loss is defined as follows:

$$los{s}_{\mathrm{cls}}=-{\displaystyle \frac{1}{N_{\mathrm{obj}}}}{\displaystyle \sum _{i=1}^{N_{\mathrm{obj}}}{\displaystyle \sum _{c\in \mathrm{classes}}\left[y_{i,c}\log \left({\widehat{y}}_{i,c}\right)+\left(1-y_{i,c}\right)\log \left(1-{\widehat{y}}_{i,c}\right)\right]}}$$

(1)

where $N_{\mathrm{obj}}$ is the number of anchor boxes containing a target, $y_{i,c}$ is the ground-truth class label, and ${\widehat{y}}_{i,c}$ is the corresponding predicted class probability. The bounding-box term measures the difference between the detected (DT) boxes and ground-truth (GT) boxes.

For anchor boxes containing a target, the detector predicts the box coordinates and applies the IoU loss to measure disagreement between the DT and GT boxes. The bounding-box loss is expressed as follows:

$$\begin{array}{l}los{s}_{\mathrm{box}}={\displaystyle \frac{1}{N_{\mathrm{obj}}}}{\displaystyle \sum _{i=1}^{N_{\mathrm{obj}}}{\displaystyle \sum _{j\in \mathrm{anchors}}{y}_{i,j}^{obj}\left[1-\mathrm{IoU}\left({\mathrm{B}}_{i\_\mathrm{dt}}, {\mathrm{B}}_{i\_\mathrm{gt}}\right)\right]}}\\ \mathrm{IoU}\left({\mathrm{B}}_{\mathrm{dt}}, {\mathrm{B}}_{\mathrm{gt}}\right)={\displaystyle \frac{\left|{\mathrm{B}}_{i\_\mathrm{dt}}\cap {\mathrm{B}}_{i\_\mathrm{gt}}\right|}{\left|{\mathrm{B}}_{i\_\mathrm{dt}}\cup {\mathrm{B}}_{i\_\mathrm{gt}}\right|}}\end{array}$$

(2)

where $y_{i,j}^{obj}$ indicates whether the ith anchor box contains a target, ${\mathrm{B}}_{i\_\mathrm{gt}}$ denotes the GT box, and ${\mathrm{B}}_{i\_\mathrm{dt}}$ denotes the predicted DT box. The confidence term compares the predicted probability of target presence with the corresponding ground-truth value using binary cross-entropy:

$$los{s}_{\mathrm{conf}}=-{\displaystyle \frac{1}{N_{\mathrm{obj}}}}{\displaystyle \sum _{i=1}^{N_{\mathrm{obj}}}{\displaystyle \sum _{j\in \mathrm{anchors}}{y}_{i,j}^{obj}\left[{\lambda }_{\mathrm{conf}}\cdot y_{i,j}^{obj}\log \left({\widehat{y}}_{i,j}^{obj}\right)+\left(1-y_{i,j}^{obj}\right)\log \left(1-{\widehat{y}}_{i,j}^{obj}\right)\right]}}$$

(3)

where ${\widehat{y}}_{i,j}^{obj}$ denotes the predicted object confidence and ${\lambda }_{\mathrm{conf}}$ is the weighting coefficient. The classification, bounding-box, and confidence losses from the P3, P4, and P5 prediction heads are weighted and aggregated to obtain $L_{\mathrm{cls}}$, $L_{\mathrm{box}}$, and $L_{\mathrm{conf}}$, respectively:

$$\left\{\begin{array}{l}{L}_{\mathrm{cls}}={\alpha }_1\cdot los{s}_{\text{cls80}}+{\alpha }_2\cdot los{s}_{\text{cls40}}+{\alpha }_3\cdot los{s}_{\text{cls20}}\\ L_{\mathrm{box}}={\alpha }_1\cdot los{s}_{\text{box80}}+{\alpha }_2\cdot los{s}_{\text{box40}}+{\alpha }_3\cdot los{s}_{\text{box20}}\\ L_{\mathrm{conf}}={\alpha }_1\cdot los{s}_{\text{conf80}}+{\alpha }_2\cdot los{s}_{\text{conf40}}+{\alpha }_3\cdot los{s}_{\text{conf20}}\end{array}\right.$$

(4)

where ${\alpha }_1, {\alpha }_2, {\alpha }_3$ denotes the weights assigned to the losses from the different prediction scales. The total objective is the weighted sum of the classification, bounding-box, and confidence terms:

$$L_{\mathrm{total}}=a{L}_{\mathrm{cls}}+b{L}_{\mathrm{box}}+c{L}_{\mathrm{conf}}$$

(5)

The confidence term is assigned the largest weight in the reported configuration. The three loss weights are set as follows: $a=0.3, b=0.3, c=0.4$.

2.2. Dataset and Training-Image Appearance Diversification

To broaden the range of visual appearances represented during training, this study uses CycleGAN [36] and a synthetic-fog procedure [26] to derive additional seasonal and fog-like training images. These transformations are used only for training-image appearance diversification. The synthetic-fog procedure is based on the dark-channel prior [37]. For a fog-free image I, the dark channel $J^{\mathrm{dark}}$ at pixel (x, y) is computed over a local patch as follows:

$$J^{\mathrm{dark}}\left(x, y\right)=\underset{c\in r,g,b}{\min }\left(\underset{\left(i,j\right)\in \omega \left(x,y\right)}{\min }\left(I^c(i,j)\right)\right),$$

(6)

where $\omega \left(x,y\right)$ denotes the local patch centered at (x, y). Atmospheric light A is estimated from the pixel with the highest dark-channel intensity. A white mask is then applied to increase the minimum values of the three color channels and produce a fog-like appearance. The resulting image is defined by:

$$\left\{\begin{array}{l}{I}_{\mathrm{fog}}\left(x, y\right)=I\left(x, y\right)\times t\left(x, y\right)+A\times \left(1-t\left(x, y\right)\right)\\ t\left(x, y\right)=e^{\left[-D\left(-0.04\sqrt{{\left(\omega -{\omega }_c\right)}^2+{\left(h-h_c\right)}^2}+s\right)\right]}\end{array}\right.,$$

(7)

where I(x, y) is the original fog-free image, A is the atmospheric-light brightness, and $t\left(x, y\right)$ is the transmission map. $\left({\omega }_c, h_c\right)$ denotes the image center, s is the fogging size, and D is the fog-thickness coefficient.

CycleGAN is also used for unpaired seasonal appearance transfer from summer insulator images to winter-like scenes with snow cover, together with reduced image resolution. The model learns mappings between two image domains while enforcing cycle consistency through two generators and their associated discriminators. A detailed schematic of the workflow is provided in Figure A2. The two generators, $G:X\to Y$ and $F:Y\to X$, map images between domains X and Y, while discriminators DX and DY distinguish real domain images from generated images. Cycle consistency requires an image translated from one domain to the other and back again to remain close to its input, represented by $F\left(G\left(x\right)\right)\approx x$ and $G\left(F\left(y\right)\right)\approx y$. The adversarial terms encourage domain-consistent outputs, and the cycle-consistency term limits content changes during the two mappings. Generators and discriminators are updated alternately during training.

The dataset was compiled from publicly available insulator images [38,39], comprising 2536 source images in total. These images were first partitioned at the source-image level into 1775 training images, 507 validation images, and 254 test images. Appearance transformations were applied only after this partitioning step. For each training image, one CycleGAN-derived image and one synthetically fogged image were generated. The final training set therefore contained 1775 original images, 1775 CycleGAN-derived images, and 1775 synthetic-fog images, corresponding to a 1:1:1 ratio and a total of 5325 training images. The validation and test sets were kept unchanged throughout all experiments, preventing transformed variants derived from training images from entering the evaluation subsets. The cumulative object-instance counts at different stages of training-image diversification are reported in

Table 1. Cumulative object-instance counts across training-image diversification stages.

Dataset Composition	Original	CycleGAN	Synthetic Fog
Insulator string	3125	5626	8127
Broken shell	1428	2561	3694
Flashover damage shell	2564	4608	6652
Total	7117	12,795	18,473

Note: Values denote cumulative annotated object-instance counts rather than numbers of images. Appearance diversification was applied only to the training subset.

, with representative examples shown in Figure 2.

Unless otherwise specified, all subsequent comparisons of localization losses, neck architectures, and backbones were conducted using this same appearance-diversified training set. Accordingly, the common PANet baseline corresponds to the appearance-diversified setting in

Table 2. Effect of training-image appearance diversification on model performance; model: CSPDarknet53 + PANet. CSPDarknet53, cross-stage partial Darknet53; P, precision; R, recall; AP, average precision; mAP, mean average precision.

Augmentation Method	Objects	P	R	AP@0.5	AP@0.5:0.95	mAP@0.5	mAP@0.5:0.95
General method	Insulator string	98.8	93.6	98.7	84.7	95.6	79.7
	Broken shell	97.3	85.4	90.5	70.7
	Flashover damage shell	93.1	96.8	97.7	83.7
CycleGAN + Synthetic fog	Insulator string	98.6	95.5	98.6	87.1	95.8	82.2
	Broken shell	98.9	81.2	90.2	71.6
	Flashover damage shell	98.3	96.8	98.7	87.9

, which achieves 95.8% mAP@0.5 and 82.2% mAP@0.5:0.95, rather than the general-augmentation setting (95.6% and 79.7%, respectively). The reported PRA and BRA gains are therefore calculated relative to PANet trained on the identical training set.

2.3. Detection Framework and Receptive-Field-Aware Adaptive Fusion

The experimental pipeline in Figure 3 places the proposed necks within a common detection framework. Training-image appearance transformations, backbone representation, cross-scale fusion, and box-regression loss are evaluated as connected design choices rather than independent research threads. PRA and BRA form the main architectural contribution; the candidate convolutional and vision-transformer backbones and the IoU-based losses are used to examine their compatibility within the same one-stage detector. This arrangement allows the experiments to relate each choice to the joint recognition of complete insulator strings and the more localized broken-shell and flashover-damage classes.

PANet [40], with its detailed topology provided in Figure A3, provides the reference fusion topology. It combines P3 (80 × 80), P4 (40 × 40), and P5 (20 × 20) features through top-down and bottom-up paths. The top-down path upsamples the low-resolution, semantically rich P5 features and merges them with higher-resolution features, after which the bottom-up path propagates spatial detail toward the lower-resolution levels. The resulting P3, P4, and P5 maps are supplied to the three detection heads. Adaptive spatial feature fusion (ASFF) [41] assigns scale-dependent weights to aligned feature maps at each pyramid level, with its detailed P5 weighting schematic provided in Figure A4. The learned weights determine the contribution of each resized feature map during fusion. Bidirectional feature pyramid network (BiFPN) [42], detailed in Figure A5, combines top-down and bottom-up propagation with an additional same-scale connection and normalized learnable weights. Together, these designs provide two complementary mechanisms for controlling information exchange across pyramid levels.

The receptive field block (RFB) [43] expands contextual coverage within the neck. Its parallel branches use different convolutional and dilation settings before their outputs are merged with the input feature map, as detailed in Figure A6. Adding RFB to PANet and BiFPN produces PANet-RFB and BiFPN-RFB (Figure A7). ASFF is then applied at the outer fusion stage to weight the aligned feature maps. The two resulting necks, PRA and BRA, therefore coordinate path aggregation, receptive-field expansion, and adaptive scale weighting within a single fusion structure. Their complete task-oriented architectures are shown in Figure 4.

2.4. Comparative Backbone and Localization-Loss Configurations

The vision transformer (ViT) family uses multi-head self-attention to model long-range interactions [44]. Swin Transformer-Tiny (Swin-T) [45], Next-ViT [46], and EfficientViT [47] are evaluated as alternative backbones to CSPDarknet53. Recent comparative studies on insulator condition monitoring have also reported strong performance from ViT-based representations relative to conventional convolutional architectures, supporting their continued evaluation for power-line visual inspection tasks [48]. The comparison examines whether their pretrained representations provide useful multiscale features for the same insulator-detection framework while the proposed contribution remains in the fusion neck. ConvNeXt V1 [49] and ConvNeXt V2 [50] provide convolutional alternatives. For each candidate, pretrained weights initialize the feature extractor before fine-tuning on the insulator dataset, while the neck and detection head are trained for the target task.

The standard IoU loss provides limited guidance in two relevant cases. If the DT and GT boxes do not overlap, it cannot represent the distance between their centers. If one fixed-size box is contained within the other, IoU remains unchanged as the inner box moves and therefore does not distinguish different internal positions. These limitations motivate the comparison of localization objectives with additional geometric terms. The evaluated alternatives account for box overlap, center distance, or aspect ratio in different ways: generalized intersection over union (GIoU) [51], distance intersection over union (DIoU) and complete intersection over union (CIoU) [52], efficient intersection over union (EIoU) [53], wise intersection over union (WIoU) [54], and the Alpha-IoU variants Alpha-GIoU, Alpha-DIoU, and Alpha-CIoU [55]. Each loss is examined within the same insulator-string and defect-detection task.

To isolate the contribution of each model component, the experiments use controlled comparisons in which only one component is varied at a time. Localization losses are compared with the backbone and neck fixed, neck architectures are compared with the backbone and localization loss fixed, and backbone alternatives are compared with the neck and localization loss fixed. The subsequent backbone-neck-loss combinations are evaluated separately to examine component compatibility and should not be interpreted as the independent contribution of any single component.

3. Results

3.1. Effect of Training-Image Appearance Diversification

The effect of extending training-image appearance is examined with a fixed CSPDarknet53-PANet detector (Table 2). General augmentation gives 95.6% mAP@0.5 and 79.7% mAP@0.5:0.95, whereas the inclusion of CycleGAN-derived and synthetically fogged images gives 95.8% and 82.2%, respectively. The greater change in the stricter metric is consistent with more accurate box localization in this comparison. Insulator strings retain higher precision (P) and recall (R) than the two defect classes in both settings, and broken-shell defects remain the most difficult. For that class, average precision (AP) at IoU thresholds from 0.5 to 0.95 increases from 70.7% to 71.6%, although recall decreases from 85.4% to 81.2%, showing that the appearance transformations do not improve all class-level measures uniformly.

3.2. Localization-Loss Comparison

To examine the influence of box regression independently of the fusion topology,

Table 3. Comparison of IoU-Loss variants; model: CSPDarknet53 + PANet. IoU, intersection over union; GIoU, generalized IoU; DIoU, distance IoU; CIoU, complete IoU; EIoU, efficient IoU; WIoU, wise IoU.

IoU Loss	Objects	P	R	AP@0.5	AP@0.5:0.95	mAP@0.5	mAP@0.5:0.95
GIoU	Insulator string	98.4	95.8	98.5	84.4	95.2	79.0
	Broken shell	95.2	81.9	89.2	69.5
	Flashover damage shell	92.0	96.4	97.7	83.3
DIoU	Insulator string	97.7	94.4	98.5	86.9	95.3	81.8
	Broken shell	96.2	77.8	88.8	71.6
	Flashover damage shell	97.4	98	98.5	86.9
EIoU	Insulator string	97.7	96.1	99.0	88.6	95.3	82.8
	Broken shell	95.9	81.1	89.0	72.1
	Flashover damage shell	94.6	95.6	98.0	87.6
WIoU	Insulator string	97.4	96.8	98.8	87.6	95.5	82.1
	Broken shell	96.8	83.7	89.1	72.5
	Flashover damage shell	94.4	98.0	98.6	86.2
CIoU	Insulator string	98.6	95.5	98.6	87.1	95.8	82.2
	Broken shell	98.9	81.2	90.2	71.6
	Flashover damage shell	98.3	96.8	98.7	87.9
Alpha-GIoU	Insulator string	97.4	96.6	99.0	86.3	95.4	80.1
	Broken shell	95.9	81.7	91.6	70.7
	Flashover damage shell	87.7	90.7	95.5	83.3
Alpha-DIoU	Insulator string	96.8	96.8	98.8	89.1	95.7	82.9
	Broken shell	97.6	83.1	89.8	71.5
	Flashover damage shell	95.0	98.3	98.3	88.0
Alpha-CIoU	Insulator string	97.6	96.8	98.8	88.9	95.9	83.1
	Broken shell	95.9	81.8	90.4	72.1
	Flashover damage shell	95.0	97.6	98.6	88.3

compares the IoU-based losses with the backbone and neck held fixed. No alternative improves every class-wise precision and recall value, and broken-shell detection remains the most difficult case throughout the comparison. On the fixed evaluation split, Alpha-CIoU yields 83.1% mAP@0.5:0.95, which is numerically the highest value among the evaluated losses. Relative to CIoU under the same CSPDarknet53-PANet configuration, Alpha-CIoU changes mAP@0.5 from 95.8% to 95.9% (+0.1 percentage points) and mAP@0.5:0.95 from 82.2% to 83.1% (+0.9 points). Because these results are obtained from a single train-validation-test partition without repeated-run variance estimation or statistical significance testing, the small differences among several IoU-loss variants should be interpreted descriptively rather than as statistically significant performance margins. CIoU and Alpha-CIoU are therefore retained as representative localization objectives for the subsequent architecture comparisons.

3.3. Cross-Scale Fusion Neck Comparison

The neck comparison uses the same appearance-diversified training set as the augmented-data setting in Table 2, while CSPDarknet53 and CIoU are held fixed (

Table 4. Comparative and ablation results for different neck architectures; backbone: CSPDarknet53; loss: CIoU. GFLOPs, giga floating-point operations; FPS, frames per second.

Necks	Model Performance Evaluation Metrics							Model Training and Inference Metrics
	Objects	P	R	AP@0.5	AP@0.5:0.95	mAP@0.5	mAP@0.5:0.95	Parameters	GFLOPs	FPS
PANet (baseline)	Insulator string	98.6	95.5	98.6	87.1	95.8	82.2	7.02M	15.8	95.8
	Broken shell	98.9	81.2	90.2	71.6
	Flashover damage shell	98.3	96.8	98.7	87.9
ASFF	Insulator string	97.7	96.6	98.9	87	94.5−1.3	79.3−2.9	9.6M	18.8	133.3
	Broken shell	92.3	79.2	87.0	67.6
	Flashover damage shell	95.5	97.6	97.6	83.4
BiFPN	Insulator string	96.0	97.5	99.1	89.0	97.1+1.3	83.3+1.1	7.17M	16.4	85.5
	Broken shell	97.5	85.4	93.1	74.0
	Flashover damage shell	94.5	98.0	98.9	86.8
PANet-RFB	Insulator string	98.4	96.7	98.9	89.8	97.2+1.4	84.9+2.7	8.75M	19.0	89.3
	Broken shell	96.7	81.8	93.5	76.0
	Flashover damage shell	98.1	98.0	99.1	88.8
BiFPN-RFB	Insulator string	97.7	95.9	98.8	88.3	97.4+1.6	84.3+2.1	8.81M	19.2	84.7
	Broken shell	97.7	87.7	94.6	76.5
	Flashover damage shell	98	98	98.7	88
Tri-BiFPN	Insulator string	98.4	95	98.6	84.3	95.2−0.6	84.2+2.0	10.12M	22.1	76.9
	Broken shell	97.5	82.4	89.2	74.0
	Flashover damage shell	96.3	96.4	97.8	89.2
PANet-RFB-ASFF (PRA)	Insulator string	98.3	96.5	99.2	90.2	98.0+2.2	86.8+4.6	14.2M	27.3	73.0
	Broken shell	100	86.6	95.3	79.6
	Flashover damage shell	98.2	99.2	99.4	90.6
BiFPN-RFB-ASFF (BRA)	Insulator string	97.4	95.8	98.6	90.4	97.3+1.5	86.2+4.0	14.2M	27.5	70.4
	Broken shell	96.2	87.8	94.6	78.0
	Flashover damage shell	95.2	98.8	98.6	90.4

). The standard YOLOv5 detection head is also kept unchanged across all neck configurations. The corresponding validation-performance curves are shown in Figure 5. Therefore, the reported PRA/BRA gains are calculated relative to PANet trained on the identical appearance-diversified training set. Relative to this common PANet reference, PRA increases mAP@0.5 by 2.2 percentage points and mAP@0.5:0.95 by 4.6 points, whereas BRA produces gains of 1.5 and 4.0 points, respectively. Tri-BiFPN and ASFF do not match PRA or BRA in this comparison, so adding fusion paths alone does not yield the same pattern. Under these controlled conditions, the larger gains at stricter IoU thresholds are observed for configurations that coordinate path aggregation, receptive-field expansion, and adaptive scale weighting rather than for those that increase path density alone. Model complexity and throughput are reported in giga floating-point operations (GFLOPs) and frames per second (FPS), respectively.

Broken-shell recall and AP remain below the corresponding values for insulator strings and flashover-damaged shells across most neck configurations, as shown in Figure 6. Most hybrid necks, except ASFF and Tri-BiFPN, improve broken-shell detection relative to their reference configurations, but the class-wise gap persists. This pattern makes it necessary to consider aggregate mAP together with class-specific recall and performance at stricter IoU thresholds.

3.4. Backbone-Neck Compatibility and Inference Efficiency

The combined backbone, neck, and localization-loss results are reported in

Table 5. ViT-based configurations were trained on a Tesla V100 (NVIDIA Corporation, Santa Clara, CA, USA), and FPS was measured on an RTX 3080 Ti (NVIDIA Corporation, Santa Clara, CA, USA).

Model Architecture		Model Performance Evaluation Metrics							Model Training and Inference Metrics
Backbone	Neck + Loss	Objects	P	R	AP@0.5	AP@0.5:0.95	mAP@0.5	mAP@0.5:0.95	Parameters	GFLOPs	FPS
Baseline model		Insulator string	98.6	95.5	98.6	87.1	95.8	82.2	7.02M	15.8	95.8
		Broken shell	98.9	81.2	90.2	71.6
		Flashover damage shell	98.3	96.8	98.7	87.9
ConvNeXt V1	PANet + CIoU	Insulator string	97.0	97.5	99.1	90.9	98.5+2.7	89.0+6.8	29.3M	74.9	70.6
		Broken shell	99.9	93.1	97.0	83.7
		Flashover damage shell	98.4	99.4	99.4	92.5
ConvNeXt V2	PANet + CIoU	Insulator string	97.7	95.9	98.4	90.6	96.6+0.8	86.4+4.2	32.74M	79.3	62.9
		Broken shell	99.9	84.0	92.3	77.5
		Flashover damage shell	98.3	98.4	99.0	91.2
Next-ViT	PANet + CIoU	Insulator string	98.1	97.9	99.3	93.0	98.8+3.0	89.9+7.7	61.49M	182.5	28.4
		Broken shell	99.9	94.4	98.1	84.6
		Flashover damage shell	98.8	99.9	99.0	92.2
EfficientViT	PANet + CIoU	Insulator string	98.4	98.0	99.3	91.1	99.0+3.2	88.2+6.0	16.01M	26.7	64.1
		Broken shell	97.9	97.9	98.9	83.4
		Flashover damage shell	98.0	98.0	99.0	90.0
Swin-T	PANet + CIoU	Insulator string	97.5	98.2	99.2	91.6	99.1+3.3	89.2+7.0	31.63M	80.0	57.1
		Broken shell	97.8	94.8	98.7	85.7
		Flashover damage shell	99.6	100	99.5	90.2
EfficientViT	PRA + CIoU	Insulator string	98.4	97.1	99.0	92.2	98.8+3.0	90.8+8.6	25.6M	42.4	45.2
		Broken shell	98.6	94.9	98.1	86.6
		Flashover damage shell	99.2	99.6	99.2	93.6
EfficientViT	BRA + CIoU	Insulator string	98.5	95.5	99.2	92.1	98.8+3.0	90.6+8.4	25.6M	42.5	41.2
		Broken shell	97.7	94.4	97.8	85.5
		Flashover damage shell	98.8	99.6	99.5	94.3
Swin-T	PRA + CIoU	Insulator string	98.1	97.0	99.2	92.8	99.1+3.3	90.5+8.3	38.8M	91.5	48.5
		Broken shell	97.9	97.1	98.7	85.8
		Flashover damage shell	99.6	100	99.5	93.0
Swin-T	BRA + CIoU	Insulator string	98.4	97.5	99.1	92.0	98.9+3.1	90.9+8.7	38.9M	91.9	48.3
		Broken shell	98.3	96.5	98.2	86.5
		Flashover damage shell	99.1	100	99.5	94.1
Swin-T	$\mathrm{PRA}+\mathsf{\alpha }-\mathrm{CIoU}$	Insulator string	98.4	96.0	99.3	93.3	98.6+2.8	91.3+9.1	38.8M	91.5	48.5
		Broken shell	97.0	92.4	97.0	86.1
		Flashover damage shell	99.7	100	99.5	94.6
Swin-T	$\mathrm{BRA}+\mathsf{\alpha }-\mathrm{CIoU}$	Insulator string	98.1	97.3	99.3	93.8	98.5+2.7	91.3+9.1	38.9M	91.9	48.3
		Broken shell	95.1	94.4	96.9	85.1
		Flashover damage shell	99.4	100	99.5	94.9

. With PANet and CIoU fixed, Swin-T achieves the highest mAP@0.5 (99.1%), while Next-ViT achieves the highest mAP@0.5:0.95 (89.9%) among the backbone substitutions. Differences between the necks are smaller at IoU = 0.5 but become clearer under mAP@0.5:0.95, as shown in Table 5 and Figure 6. Swin-T-BRA with CIoU achieves 98.9% mAP@0.5 and 90.9% mAP@0.5:0.95. Swin-T-PRA + Alpha-CIoU and Swin-T-BRA + Alpha-CIoU jointly achieve the highest mAP@0.5:0.95 of 91.3%, with corresponding mAP@0.5 values of 98.6% and 98.5%, respectively. Selecting a stronger backbone therefore does not eliminate the performance differences associated with the neck and localization loss. To make the contribution of each component explicit, the controlled neck, backbone, and localization-loss comparisons are summarized in

Table 6. Summary of controlled comparisons for training-image diversification and model components.

Component Evaluated	Fixed Configuration	Controlled Comparison	mAP@0.5	ΔmAP@0.5	mAP@0.5:0.95	ΔmAP@0.5:0.95
Training-image diversification	CSPDarknet53 + PANet	General augmentation → CycleGAN + fog	95.6 → 95.8	+0.2	79.7 → 82.2	+2.5
Neck	CSPDarknet53 + CIoU	PANet → PRA	95.8 → 98.0	+2.2	82.2 → 86.8	+4.6
Neck	CSPDarknet53 + CIoU	PANet → BRA	95.8 → 97.3	+1.5	82.2 → 86.2	+4.0
Backbone	PANet + CIoU	CSPDarknet53 → Next-ViT	95.8 → 98.8	+3.0	82.2 → 89.9	+7.7
Localization loss	CSPDarknet53 + PANet	CIoU → α-CIoU	95.8 → 95.9	+0.1	82.2 → 83.1	+0.9

Note: All neck, backbone, and localization-loss comparisons use the same appearance-diversified training set; therefore, their reported gains do not include the improvement introduced by training-image diversification itself.

. The corresponding validation-performance curves for the evaluated model architectures are shown in Figure 7.

Performance, computational complexity, and model-only inference speed are compared in Figure 8. Models without an explicit neck label use PANet. Next-ViT has the largest GFLOPs count and the lowest FPS among the reported models. Although EfficientViT-PRA and EfficientViT-BRA require fewer GFLOPs than Swin-T-BRA, their measured FPS is lower, showing that fewer operations do not translate into higher throughput on the reported hardware. The Swin-T configurations with PRA or BRA occupy similar performance-efficiency ranges in Figure 8 and Table 5.

The F1 score (F1)-confidence curves in Figure 9 show how the Swin-T configurations behave as the confidence threshold changes. For broken-shell defects, Swin-T-BRA reaches an F1 score of 0.978 at a threshold of 0.771 and remains competitive across the displayed range (Figure 9c). The curves complement the fixed-threshold metrics by showing the precision-recall trade-off within the evaluated data. Representative qualitative detection results and Score-CAM response maps for insulator strings (Figure A8), broken-shell defects (Figure A9), and flashover-damaged shell defects (Figure A10) are provided in Appendix A.

4. Discussion

The controlled comparisons summarized in Table 6 distinguish the effects of training-image appearance diversification, neck architecture, backbone, and localization loss. With CSPDarknet53 and PANet fixed, appearance diversification increases mAP@0.5 from 95.6% to 95.8% (+0.2 percentage points) and mAP@0.5:0.95 from 79.7% to 82.2% (+2.5 points). Using the same appearance-diversified training set, and with CSPDarknet53 and CIoU fixed, replacing PANet with PRA further improves mAP@0.5 and mAP@0.5:0.95 by 2.2 and 4.6 percentage points, respectively, while BRA provides additional gains of 1.5 and 4.0 points. These gains therefore represent the architectural contribution attributable specifically to PRA/BRA under the controlled setting, independent of the improvement introduced by training-image appearance diversification. In comparison, changing only the backbone from CSPDarknet53 to Next-ViT with PANet and CIoU fixed produces gains of 3.0 and 7.7 points, whereas replacing CIoU with Alpha-CIoU under the fixed CSPDarknet53-PANet configuration produces smaller gains of 0.1 and 0.9 points. Accordingly, the absolute performance of the combined backbone-neck-loss configurations in Table 5 reflects the joint effects of training-image appearance diversification and model-component design, whereas the reported PRA/BRA gains quantify the incremental architectural improvement relative to PANet trained on the identical appearance-diversified training set.

These results show that PRA/BRA improve detection performance under stricter IoU requirements when the backbone, localization loss, and detection head are fixed. However, mAP@0.5:0.95 remains an aggregate detection metric and does not independently quantify bounding-box regression error. Alternative regression-head architectures and direct geometric localization metrics, such as center-distance error, were not separately evaluated in the present study. Therefore, the observed gains are interpreted as improved high-IoU detection performance rather than as direct evidence of an isolated localization mechanism. The relatively small differences among several localization losses in Table 3 should also be interpreted as numerical differences on the present fixed data split, because repeated training runs, cross-validation, and statistical significance tests were not performed.

The configuration with the highest strict-localization performance is distinguished from the final model selected for transmission-line defect inspection. Swin-T-PRA + Alpha-CIoU and Swin-T-BRA + Alpha-CIoU both achieve the highest mAP@0.5:0.95 of 91.3%. However, recall is a critical consideration in transmission-line defect inspection because missed detections may leave defects unidentified during inspection. From this perspective, Swin-T-BRA + CIoU provides a more suitable performance trade-off, achieving 98.9% mAP@0.5 and 90.9% mAP@0.5:0.95, together with a broken-shell recall of 96.5% and a flashover-damage recall of 100%. By comparison, Swin-T-PRA + Alpha-CIoU and Swin-T-BRA + Alpha-CIoU achieve broken-shell recalls of 92.4% and 94.4%, respectively, despite their slightly higher mAP@0.5:0.95 of 91.3%. Swin-T-BRA + CIoU also achieves the highest overall F1 score of 0.983 among the evaluated Swin-T configurations while maintaining an inference speed of 48.3 FPS. Accordingly, considering the importance of reducing missed defect detections together with overall detection accuracy, F1 performance, and inference efficiency, Swin-T-BRA + CIoU is selected as the final model configuration in this study.

Broken-shell defects remain more difficult than complete insulator strings and flashover-damaged shells. This persistent gap is associated with both data imbalance and target-scale disparity. As shown in Table 1, broken-shell instances are less frequent than the other two classes, and their damaged regions generally occupy smaller image areas and depend strongly on viewing direction and local geometry. These factors can make discriminative cues more susceptible to attenuation during feature extraction and cross-scale fusion. Although PRA and BRA improve broken-shell recall and AP@0.5:0.95 relative to the PANet baseline, the class-wise gap remains. For example, PRA increases broken-shell recall from 81.2% to 86.6% and AP@0.5:0.95 from 71.6% to 79.6%, while BRA increases them to 87.8% and 78.0%, respectively. The receptive-field-aware fusion mechanisms therefore mitigate, but do not fully resolve, the difficulty associated with this class. Aggregate mAP should consequently be interpreted together with class-specific recall and AP.

Among the frameworks evaluated on the same task, the Swin-T-BRA + CIoU configuration records the highest mAP@0.5 and mAP@0.5:0.95 values reported in

Table 7. Comparative results for different detection frameworks. YOLO, You Only Look Once; SSD, single shot multibox detector; Faster R-CNN, Faster region-based convolutional neural network.

Models	Objects	Precision	Recall	AP@0.5	AP@0.5:0.95	mAP@0.5	mAP@0.5:0.95
YOLO V5	Insulator string	98.8	93.6	98.7	84.7	95.6	79.7
	Broken shell	97.3	85.4	90.5	70.7
	Flashover damage shell	93.1	96.8	97.7	83.7
YOLO V3	Insulator string	95.1	93.2	97.4	81.1	94.5	79.6
	Broken shell	95.3	82.6	88.3	71.2
	Flashover damage shell	92.3	96.8	97.8	86.6
YOLO V7	Insulator string	95.7	98.4	98.8	87.4	96.6	80.0
	Broken shell	94.5	85.4	96.2	72.0
	Flashover damage shell	86.4	88.4	94.9	80.6
YOLO V8	Insulator string	96.2	99.0	99.3	93.7	97.9	86.0
	Broken shell	94.7	86.6	95.4	76.0
	Flashover damage shell	96.4	97.2	98.9	88.4
CenterNet	Insulator string	95.6	95.6	96.3	76.8	93.8	69.7
	Broken shell	93.8	90.5	93.0	62.8
	Flashover damage shell	78.7	90.5	92.0	69.6
SSD	Insulator string	95.3	92.3	98.1	80.6	94.7	72.4
	Broken shell	94.9	76.0	94.7	67.2
	Flashover damage shell	92.1	77.0	91.9	69.2
EfficientDet	Insulator string	96.4	74.1	88.6	68.2	91.9	67.8
	Broken shell	94.0	86.5	94.0	63.4
	Flashover damage shell	83.0	87.0	93.1	71.8
Faster R-CNN	Insulator string	88.0	99.4	99.0	74.9	97.3	65.8
	Broken shell	86.3	95.2	95.9	58.7
	Flashover damage shell	72.9	97.6	97.0	63.9
Swin-T-BRA(Ours)	Insulator string	98.4	97.5	99.1	92.0	98.9	90.9
	Broken shell	98.3	96.5	98.2	86.5
	Flashover damage shell	99.1	100	99.5	94.1

. This result supports its relative performance within the dataset and evaluation setup used for the table. Because performance depends on data composition, training protocol, and implementation details, the comparison does not provide a universal ranking. Broader comparisons would require controlled evaluation on additional datasets and acquisition conditions.

5. Conclusions

This study investigated whether receptive-field expansion and adaptive cross-scale feature aggregation can improve high-IoU localization when complete insulator strings, localized broken-shell defects, and flashover-damaged regions are detected within a unified transmission-line inspection framework. To address their substantial differences in scale and contextual dependence, PRA and BRA were developed to integrate multi-path feature propagation, receptive-field refinement, and adaptive scale weighting within the detection neck.

Controlled comparisons with CSPDarknet53 and CIoU show that the proposed necks provide larger gains under stricter localization criteria than at IoU = 0.5. Relative to PANet, PRA improves mAP@0.5 and mAP@0.5:0.95 by 2.2 and 4.6 percentage points, respectively, while BRA yields gains of 1.5 and 4.0 points. Further backbone-neck-loss comparisons show that these gains depend on the interaction between feature representation, cross-scale aggregation, and boundary regression. Swin-T-PRA + Alpha-CIoU and Swin-T-BRA + Alpha-CIoU jointly achieve the highest mAP@0.5:0.95 of 91.3%. Considering the importance of recall in transmission-line defect inspection, together with overall detection performance and inference efficiency, Swin-T-BRA + CIoU is selected as the final model configuration in this study.

The results show that receptive-field-aware adaptive cross-scale fusion is particularly effective under stricter IoU-based detection criteria for heterogeneous-scale insulator targets. The present conclusions remain limited to the evaluated dataset and protocol. Future work will examine repeated-run variability and statistical significance, class-balanced and small-defect-oriented training strategies, cross-dataset and cross-line generalization, adverse-weather conditions, and end-to-end deployment on resource-constrained inspection platforms.