Target detection model based on traditional methods

In the early stages of computer vision, target detection mainly relies on the combination of manual feature extraction and traditional classifiers, which are widely used in a variety of scenes. Early studies mostly used traditional methods based on the combination of Haar features and cascade classifiers. Specifically, Haar features scan the image block by block through a sliding window and combine with a pre-trained Haar cascade classifier to accomplish classification and detection. Zhao et al. [10] proposed to retain insulators based on orientation angle detection and a priori knowledge of the insulator’s binary shape and localize them with a minimum outer rectangular frame. By traversing all possible orientation angles, multiple insulators with different orientation angles can be localized. Jaya Bharata Reddy et al. [11] proposed the use of Discrete Orthogonal S-transform to extract image features and combined with Support Vector Machine (SVM) and Adaptive Neuro Fuzzy Inference System (ANFIS) to estimate the insulator condition and combined with Support Vector Machine classification to identify the state of the insulator. Zuo et al. [12] proposed a classifier that can recognize and locate insulators obtained by feature extraction and training with Haar features, integral map, cascade classifier, and directional gradient histogram, then segment the insulators by a series of digital image processing methods, and finally the insulator pixels obtained by segmentation are statistically analyzed to determine whether the insulators are missing or not. In order to improve the robustness of the model in such environments, Schwegmann et al. [13] proposed a novel vessel detection method, which first performs an initial screening by a constant false alarm rate (FAR) pre-screening step, and then uses a cascading classifier based on Haar-like features for vessel identification. However, these methods provide a solid foundation for target detection, they require manual design of feature extractors, have limited feature expressiveness, limited algorithmic recognition accuracy, and are difficult to cope with complex scenarios.

Deep learning based target detection method

With the development of deep learning, specific network structures for target detection have emerged, which are divided into two main categories: single-stage detectors and two-stage detectors.

The single-stage detector directly predicts the targets in the image and accomplishes the detection of target location and category. Liu et al. [14] proposed Single Shot MultiBox Detector (SSD), a single-step anchor-box based target detection algorithm, which is able to predict a set of default bounding boxes from feature maps at different scales, handle targets of different sizes, and predict both the category confidence and adjustment within each bounding box. Joseph Redmon [15] proposed You Only Look Once (YOLO), which treats target detection as a regression problem and directly predicts the bounding boxes and their category probabilities in an image, providing a basis for the optimization of the subsequent YOLO series of algorithms.

The two-stage detector divides target detection into two steps: candidate region generation and classification. Ross Girshick et al. [16] proposed the RCNN, which combines region proposals with deep convolutional neural network (CNN) features. The method generates candidate regions through unsupervised learning, extracts features using pre-trained CNNs, and ultimately classifies the target with an SVM classifier. Subsequently, Ross Girshick et al. [17] proposed Fast R-CNN based on RCNN, where candidate regions are mapped to a fixed-size feature map through ROI pooling layer, and the image needs only one forward propagation to extract features, and then a fully connected layer is used to complete the classification and regression tasks, which significantly improves the speed and efficiency of the RCNN. Ren et al. [18] further proposed the Regional Proposal Network (RPN) to share the convolutional features of the image with Fast R-CNN, which reduces the computational cost.

These network structures have achieved significant results in target detection. By training deep learning models, they are able to detect classroom behaviors and improve the accuracy and efficiency of detection. However, these general methods have some limitations in different scenes and object detection, and still need a lot of adjustment and optimization for specific scenes.

Improved deep learning based target detection

The general target detection method has certain limitations, but through the improved method based on the basic algorithm, we can see its feasibility. Haiwei et al. [19] proposed an improved YOLOv8 classroom behavior detection model, which improves the effectiveness and accuracy of YOLOv8 in classroom behavior detection by combining Res2Net and YOLOv8 network modules, proposing the C2f_Res2block module, and introducing MHSA and EMA mechanisms. In addition, to further enhance the feature extraction effect of the model, Wenqi et al. [20] proposed the CSB-YOLO model, which enhances multi-scale feature fusion through bidirectional feature pyramid network (BiFPN), designs an efficient reparameterized detection head (ERD Head) to improve the inference speed, and introduces self-calibrating convolution (SCConv) to compensate for the loss of accuracy in the lightweight design.

Li et al. [21] utilized a feature pyramid network to fuse the features of the upsampling layer and the invariant scale convolutional layer, and retained the multi-scale feature layer extracted from the traditional SSD structure, which improved the accuracy of detection. Ding et al. [22] replaced IOU with EIOU to calculate the box loss and introduced Assumption-free K-MC2 (AFK-MC2) algorithm based on yolov5, and cluster non-maximum suppression (Cluster-NMS), to improve the leakage of detection due to occlusion. Henghuai et al. [23] proposed an improved YOLO-v4 behavior detection algorithm that combined a cross-stage local network and an embedded connection component to improve the recognition ability of student and teacher behaviors. In order to solve the problem of difficult recognition of student behavior in the case of occlusion, the Repulsion loss function is added to reduce the misdetection and omission through RepGT and RepBox loss. Hsu et al. [24] proposed the RSA-YOLO (Ratio and Scale Aware YOLO) method, which aims to solve the problem of poor detection performance in pedestrian detection due to the small target scale and the large difference in the aspect ratio of the input image. The method dynamically adjusts the input layer hyperparameters of YOLOv3 by introducing a scale-aware mechanism, and adopts intelligent image segmentation technology to improve the detection accuracy. Meanwhile, a scale-aware mechanism of multi-resolution fusion is proposed to effectively improve the detection of pedestrians with small targets.

Tang et al. [25] proposed a YOLOv7-based target detection method, called TOD-YOLOv7, to address the challenges of low-resolution images, dense occlusions, and different poses in the TinyPerson dataset. By adding a small target detection layer to the YOLOv7 network, introducing a recursive gated convolution module and a coordinate attention mechanism, the model’s detection capability is enhanced and inference time is reduced. In addition, data enhancement techniques are combined to improve the algorithm’s representation learning capability. Wu et al. [26] proposed a Faster R-CNN-based Different Scale Face Detector (DSFD), which aims to solve the challenge of small-scale face detection. The method first obtains face ROIs through a multi-task regional proposal network (RPN) combined with augmented face detection, and assigns the proposals to three corresponding Fast R-CNN networks based on different proposal scales. Wang et al. [27] proposed a UAV target detection model based on YOLOv8 optimization, called UAV-YOLOv8, aiming to solve the problem of low accuracy and resource constraints in detecting small targets in UAV images. The model introduces Wise-IoU (WIoU) v3 as a regression loss function to improve the localization ability and optimizes the backbone network through BiFormer attention mechanism. A Focal FasterNet module is also designed for multi-scale feature fusion to improve the detection performance of small targets.

Tianyong et al. [28] proposed an improved YOLOv8 network called YOLO-SE, which aims to address the challenges of multi-scale target detection and small target detection in remote sensing images. By introducing lightweight SEConv convolution and SEF modules, the network parameters are reduced and the detection speed is accelerated. Also, an efficient multi-scale attention (EMA) mechanism is integrated to enhance the feature extraction capability. The network also contains specialized tiny target detection heads and uses Transformer [29] prediction heads instead of the original detection heads. In addition, the Wise-IoU loss function is introduced to cope with the gradient problem for low-quality instances. Wang et al. [30] proposed an adaptive enhancement fusion framework based on YOLO (YOLO_AEF). This framework employs a multiple exposure enhancement module (MEEM) to improve image quality under complex lighting conditions and utilizes an adaptive feature fusion module (AFFM) to fuse raw and enhanced image features, thereby enhancing robustness and contextual expression. Concurrently, it incorporates a fusion detection module (FDM) to achieve robust occlusion detection.

Although generalized target detection methods have shown better results, there are still some challenges for the detection of transmission line components. For example, the detection and target overlapping of existing methods for small-sized targets such as Spacer can affect the effectiveness of the deep learning model, yet it is difficult to obtain a large and sufficient amount of data for training due to the difficulty of transmission line data acquisition. Therefore, the research in this paper focuses on solving the problems of overlapping targets and large size differences in transmission line component detection. By improving the training method of the deep learning model, the detection accuracy and robustness of transmission line components can be improved.

Improved neck module

In target detection tasks, the occurrence of occlusion phenomenon leads to the loss of local information about the target, which poses a challenge to the recognition performance of the model. The traditional Neck module, despite its self-attention mechanism, fails to effectively capture the contextual information of occluded objects due to its high computational complexity, leading to its unsatisfactory performance in dealing with occlusion situations. In order to counteract this problem, we have made improvements to the Neck module. Specifically, the ACmix module [31] was added after the output channels of the C2f and C2fCIB. With ACmix’s decomposition and reconstruction strategy, we are able to extract the contextual information of the occluded target more efficiently, Enhanced detection accuracy.

This improvement leads to a significant optimization of the model’s performance in occlusion situations. ACmix combines the advantages of convolution and self-attention mechanisms (self-attention). Through analysis, both techniques operate similarly in the first stage, relying mainly on 1×1 convolution for feature projection, with computational complexity quadratically related to the number of channels. Based on this finding, ACmix shares convolutional and self-attentive projection operations, reducing computational overhead. While keeping the computational cost low, ACmix achieves significant performance gains on tasks such as image classification and target detection. Fig 2(a) and 2(b) represent the standard convolution and self-attention mechanisms, respectively. The workflow of the ACmix model is demonstrated in Fig 2(c).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Architecture of the ACmix module.

https://doi.org/10.1371/journal.pone.0335186.g002

ACmix forms a hybrid module by combining convolution and self-attention mechanisms. In the first stage, the input feature maps are projected through three convolutions to generate multiple intermediate feature maps. In the second stage, these intermediate feature maps follow two different aggregation methods: convolutional paths and self-attention paths.

For the self-attentive path, the generated intermediate feature maps are divided into groups and each group contains three feature maps which are used as query, key and value. Through the self-attention mechanism, the similarity between the query and the key is computed and the values are weighted and summed using the attention weights. This process can be represented by the following equation:

(1)

where q_ij and k_ab are the feature representations of the query and key, respectively, denote the attention weights, and are the values.

For convolutional paths, a lightweight fully connected layer is used to generate k² feature maps, which are shifted and summed by the following formula:

(2)

Where denotes the feature map obtained after applying the convolution kernel, and is a shift operation.

Eventually, the outputs from the convolutional and self-attentive paths are weighted and summed to form the final output. The formula is given below:

(3)

where α and β are learnable scalars that control the output strength of the self-attentive path and the convolutional path, respectively.

All figures and tables should be cited in the main text as Fig 1, Table 2, etc.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Training parameter settings.

https://doi.org/10.1371/journal.pone.0335186.t002

Improved backbone and head modules

The improved Backbone module adds C2f-ELAN4 (GELAN) [32] to the output channel of CBS, GELAN incorporates CSPNet and ELAN mechanisms and utilizes RepConv to obtain more effective features while specializing in single-branching structure in inference, thus increasing the detection accuracy at multiple scales. The GELAN module improves feature extraction by combining multiple convolutional operations. As in Fig 3. first, the input feature map undergoes an initial convolution operation to generate an output feature map with channels, expressed as:

(4)

where , H and W are the height and width of the feature map, respectively, and C is the number of channels. Next, the generated feature map Y is equally divided into two parts Y₁ and Y₂ with channels each, which can be expressed as:

(5)

where , the segmented feature map Y₂ is subsequently processed through the RepNCSP module. The core of RepNCSP is a convolution operation combining and . Eq:

(6)

where, this design allows the model to simultaneously process features from different receptive fields, enhancing the ability to characterize complex scenes.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Schematic diagram of the ASFF module.

https://doi.org/10.1371/journal.pone.0335186.g003

Next, the two RepNCSP-processed features Z₂, Z₃ and Y₁ are re-fused by a splicing operation to form a complete feature map Z, denoted as:

(7)

where . Finally, the fused feature map Z is again subjected to a convolution operation to generate the final output , Eq:

(8)

The design improves the expressive power of the model in feature extraction by separating the channels, feature processing with different convolutional kernels and the final fusion convolution operation, which is particularly suitable for tasks such as target detection where multi-scale features need to be processed.

The improved Head module adds ASFF [33] to the output channels of ACmix, which introduces an adaptive spatial feature fusion method that effectively filters out conflicting information and thus enhances scale invariance. The proposed ASFF (Adaptively Spatial Feature Fusion) module aims to improve multi-scale target detection by fusing feature maps from different scales. Fig 4 illustrates the feature maps from different levels of the feature pyramid (Level 1, Level 2, Level 3) with different resolutions and strides. The ASFF module adjusts the weighting of the features at each scale using the adaptive weights , which are combined with the feature maps of each level (X1 3, X2 3, X3 3) by element-by-element multiplication, ensuring that the contributions of each level can be effectively integrated. to ensure that the feature contributions of each layer can be effectively integrated. Especially for multiscale detection, low-level high-resolution features are crucial, and the ASFF module generates unified features for subsequent target prediction by adaptively fusing different scale features, and, finally, fusing the scale features after an additive operation.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Schematic diagram of the GELAN architecture.

https://doi.org/10.1371/journal.pone.0335186.g004

Experimental setup and data description

The code for the paper was implemented on a server equipped with 2 RTX4090 GPUs. As shown in Table 3, compared to earlier datasets such as Power line dataset (2017) and Tomaszewski et al. (2018) that contain only a single asset class and lack defective samples, the InsPLAD dataset has a significant lead in terms of the number of asset classes (17), the number of annotations (28,933) and the number of images (10,607), and covers 5 classes of defective assets. significantly ahead of the others and covers 5 classes of defective assets, supporting a variety of visual tasks such as Object Detection (OD), Image Classification (IC) and Anomaly Detection (AD). Therefore, the dataset in this paper adopts InsPLAD, a dataset and benchmark for power line asset inspection, with specific parameters shown in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Public power line datasets.

https://doi.org/10.1371/journal.pone.0335186.t003

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Basic parameters of the InsPLAD dataset.

https://doi.org/10.1371/journal.pone.0335186.t004

Ablation experiment

In Table 5, we show the results of the ablation experiments of the model by systematically investigating the independent contribution of each module to the model performance in order to assess its importance in complex detection tasks. In this set of experiments, experiments 0, 1, 2, 3, 4, 5, 6 and 7 were tested on raw data. Labeled A(Acmix), B(ASFF), C(GELAN), where a0, a1, a2 are for multi-scale target detection and a3, a4, a5 are for occluded target detection.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Ablation study results of YOLOv10 with different modules.

https://doi.org/10.1371/journal.pone.0335186.t005

First, in the base model without any improvement module (experiment a), the overall mAP@50 performance is 0.838. Due to the lack of specific module support, the base model has obvious deficiencies in coping with target occlusion and scale differences. For example, the low detection effect of Stockbridge Damper and Yoke Suspension indicates that the base model is limited in dealing with these complex detection tasks as shown by a(0) in Fig 2. The prediction of Stockbridge Damper reaches 0.7 and the prediction of Yoke Suspension reaches 0.45.

The experiment with the addition of the Acmix module (experiment d) shows that the detection effect is as shown in Fig 5, b(5), where the Polymer Insulator Tower Shackle detection efficiency is improved from 0.3 to 0.91 compared to the original a(5). b(4) in Fig 5 also shows that compared to the original a(4), the Polymer Insulator Tower Shackle detection efficiency is improved from 0.86 to 0.89 compared with the original a(4) in Fig 5, which effectively solves the target occlusion problem.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Detection results of YOLOv10 with progressively enhanced modules.

https://doi.org/10.1371/journal.pone.0335186.g005

The experiment with the addition of GELAN and ASFF modules (experiment e), the detection effect is shown as b(1) in Fig 5, compared with the original a(1) Yoke Suspension, the detection effect is improved to 0.84 improved to 0.91, and Stockbridge Damper is improved from 0.41 to 0.89, respectively. improves the multi-scale complex background detection efficiency and effectively enhance multi-scale target detection.

Finally, the complete model containing all enhancement modules (experiment h) achieves the best performance on all categories, overall mAP@50 raised to 0.950. The synergy of the modules with the introduction of data enhancement strategies not only improves the detection performance of the model in multi-scale, complex background and occluded target environments, but also significantly enhances the model’s generalization ability.

Comparison experiment

In order to verify the soundness of the proposed method, we compare it with several classical target detection methods, yolo11 [34], rtdetr-l [35], yolov8 [36], rtdetr-x [37], yolov9 [38], rtdetr-resnet50 [39], yolov10 [40], rtdetr-resnet101 [41]. As shown in Table 6. Our improved model ACC_yolov10_Enhance (ours) addresses the two key challenges of the transmission equipment inspection task (overlapping targets and large size difference) and significantly outperforms the other models in mAP@50. With the tabular data, we can clearly see how ACC_yolov10_Enhance (ours) compares with other models on different categories, further validating its advantages in complex detection tasks.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Performance comparison of object detection models for transmission equipment inspection (mAP@50).

https://doi.org/10.1371/journal.pone.0335186.t006

In the “target occlusion” problem, ACC_yolov10_Enhance (ours) utilizes the Acmix module to combine the self-attention mechanism with convolutional feature extraction to improve the model’s discriminative ability in complex backgrounds. In the category of “lightning rod shackle”, the mAP@50 of ACC_yolov10_Enhance (ours) reaches 0.924, which is significantly higher than that of yolo11 (0.794) and yolov8 (0.82), and the mAP@50 of ACC_yolov10_Enhance (ours) is 0.924, which is significantly higher than that of yolo11 (0.794) and yolov8 (0.82). In the category of “Polymer Insulator Tower Shackle”, d(3) 0.75 in Fig 6 is improved to c(3) 0.8. The results show that the improved model has stronger detection accuracy in the scene with more serious target occlusion.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Comparative visualization of detection performance across different models (Ours vs. YOLOv8/RT-DETR).

https://doi.org/10.1371/journal.pone.0335186.g006

To address the problem of “large size difference”, ACC_yolov10_Enhance (ours) enhances the adaptability to multi-scale targets through the GELAN and ASFF modules. For example, in the category of “polymer insulator lower shackle”, which has a large size difference, although other models such as yolo11 and yolov9 also have high performance in this category (0.907 and 0.928, respectively), ACC_yolov10_Enhance (ours) is not able to adapt to multiscale targets through the GELAN and ASFF modules. yolov10_Enhance (ours) is more robust in the overall test, and in particular more balanced in terms of overall performance in the different categories. In addition, in the “yoke” category, ACC_yolov10_Enhance (ours) reaches 0.973, while yolov10 only reaches 0.953. In the “Yoke Suspension, Polymer Insulator Tower Shackle” category, ACC_yolov10_Enhance (ours) is more robust in the overall test, especially in the balanced performance of the different categories. Polymer Insulator Tower Shackle,” in the category of “Yoke Suspension,” d(5) 0.47,0.81 in Fig 6 improves to c(5) 0.95,0.88. This shows the excellent performance of the improved model on large size targets.

Combining the results of “all classes”, the overall mAP@50 of ACC_yolov10_Enhance (ours) reaches 0.95, which is significantly higher than that of yolov9, which is the best performer among other models (0.901). In the overall comparison of different categories, ACC_yolov10_Enhance (ours) shows high stability and accuracy under all kinds of challenges, providing higher reliability and accuracy for the automated inspection of transmission equipment.

Figs 7, 8, and 9 demonstrate the trends of key loss terms and evaluation metrics for the ACC_yolov10_Enhance model over 150 training cycles. Specifically, the bounding box loss (box loss), classification loss (cls loss), and distribution focus loss (DFL loss) on the training and validation sets decrease rapidly at the initial stage, and then level off, indicating that the model converges gradually. In addition, the performance metrics such as precision, recall, and average precision (mAP@50 and mAP@50-95) increase rapidly in the early stage and stabilize at a high level in the later stage, reflecting the gradual improvement of the model’s performance on the detection task and its good generalization ability. Overall, the figure shows that the model achieves good convergence during training and stable performance on the validation set.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Convergence trends of training loss components (Box/Cls/DFL).

https://doi.org/10.1371/journal.pone.0335186.g007

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Evolution of model performance metrics during training (Precision/Recall/mAP).

https://doi.org/10.1371/journal.pone.0335186.g008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Validation loss analysis across training epochs (Box/Cls/DFL).

https://doi.org/10.1371/journal.pone.0335186.g009