A Two-Stage Corrosion Defect Detection Method for Substation Equipment Based on Object Detection and Semantic Segmentation

Abstract

Corrosion defects will increase the risk of power equipment failure, which will directly affect the stable operation of power systems. Although existing methods can detect the corrosion of equipment, these methods are often poor in real-time. This study presents a two-stage detection approach that combines YOLOv8 and DDRNet to achieve real-time and precise corrosion area localization. In the first stage, the YOLOv8 network is used to identify and locate substation equipment, and the detected ROI areas are passed to the DDRNet network in the second stage for semantic segmentation. To enhance the performance of both YOLOv8 and DDRNet, a multi-head attention block is integrated into their algorithms. Additionally, to address the challenge posed by the scarcity of corrosion defect samples, this study augmented the dataset using the cut-copy-paste method. Experimental results indicate that the improved YOLOv8 and DDRNet, incorporating the multi-head attention block, boost the mAP and mIoU by 5.8 and 9.7, respectively, when compared to the original method on our self-built dataset. These findings also validate the effectiveness of our data augmentation technique in enhancing the model’s detection accuracy for corrosion categories. Ultimately, the effectiveness of the proposed two-stage detection method in the real-time detection of substation equipment corrosion defects is verified, and it is 48.7% faster than the one-stage method.

1. Introduction

With the increasing number of substations, ensuring the continuous and stable operation of substation equipment has become the key to maintaining the stability of the power system. These devices are exposed to the changing climate and environment for a long time, and they are affected by humidity, temperature changes, rainfall, and pollutants, which will gradually corrode, thereby reducing the performance and reliability of the equipment. Currently, the supervision and maintenance of substations mainly rely on manual inspection, although this process has been relatively perfect, but there are still problems of low efficiency and high cost. Therefore, there is an urgent need to develop an intelligent, automated corrosion area detection technology to improve inspection efficiency and quality, as well as ensure the smooth operation of the power system.

Currently, research on the corrosion detection of substation equipment mainly focuses on the corrosion detection of grounding networks [1,2,3]. Common methods include the ultrasonic method [4], electromagnetic method [5], and electrochemical method [6]. However, there are fewer studies on the corrosion detection of other equipment in substations (e.g., transformers, capacitors, mutual inductors, disconnect switches and circuit breakers). Therefore, this paper aims to explore a corrosion detection algorithm with wider applicability. In recent years, machine vision inspection algorithms have made significant progress in the field of surface defect detection [7,8,9,10,11,12] and have received widespread attention due to their non-contact detection means and low cost. Therefore, this paper attempts to solve the problem of detecting corroded areas in substations using vision sensor-based artificial intelligence algorithms. Object detection [7,13] and semantic segmentation [14,15] are the two most common types of algorithms in the field of machine vision.

Object detection algorithms are now widely used in industrial production, which can be divided into two categories: one-stage networks and two-stage networks. The representative networks of two-stage networks are the RCNN series [16,17]. The task of two-stage networks is usually divided into two steps: first using the region proposal algorithm to generate candidate regions that may contain objects and then carrying out accurate coordinate regression and classification of these candidate regions. In contrast, one-stage object detection networks directly perform dense prediction of the entire image without an intermediate step of generating candidate regions, thus allowing direct determination of the location, size, and class of the detected target. Representative models of such networks are the YOLO series [18,19,20] and the SSD series [21], among others. Since one-stage object detection networks do not need to generate candidate regions, detection time is saved. This method can achieve real-time detection while ensuring detection accuracy, thus becoming an important research direction for the current defect detection of substation equipment. In this paper, the latest YOLOv8 network is selected as the object detection network.

Semantic segmentation is able to classify each pixel in an image and therefore accurately identify the boundaries and shape of corroded areas. This is important for determining the extent and location of corrosion on metal surfaces. The current mainstream semantic segmentation algorithms can be categorized as encoder–decoder structure, image pyramid structure, and two-branch structure. The encoder–decoder structure is represented by U-Net [22], which tends to be symmetric, connects the encoder and decoder parts through jump connections to improve the ability to capture details, and is widely used in medical image segmentation. The image-pyramid-based structure is represented by the DeepLab family [23,24], which uses dilated convolution and null-space pyramid pooling to capture contextual information at multiple scales, thus improving the accuracy of segmentation. Semantic segmentation algorithms with a two-branch structure, represented by DDRNet [25], usually include a main branch for extracting high-level semantic information, which focuses on global contextual information. And a secondary branch for extracting detailed features is used to ensure that the detailed parts of the segmentation results are accurately processed. The two-branch design allows the network to balance segmentation accuracy and real-time performance, which is suitable for scenarios that require fast processing. The timely detection and treatment of corrosion problems can avoid further deterioration of the corrosion problem, thus reducing the cost of large-scale repair or replacement.

Although the two-branch semantic segmentation network already has good real-time performance, the substation monitoring image contains a lot of noise information other than the equipment to be detected, such as the sky, the road, some equipment that does not corrode, etc. In order to solve the problem, this paper tries to use a two-stage detection method. Detecting these regions will consume a certain amount of time, which affects the real-time performance of the detection system. In order to solve this problem, this paper tries to adopt a two-stage detection method, i.e., firstly, we use the target detection algorithm to find out the region of interest (ROI) in the monitoring image and the pixel regions of the equipment to be detected, and then we feed the ROI into the semantic segmentation network for local detection. Since the computational complexity of the object detection algorithm is much lower than that of the semantic segmentation algorithm, the two-stage detection method will save a lot of time compared to the pixel-by-pixel detection of the whole image using only the semantic segmentation algorithm. The object detection and semantic segmentation networks used in this paper are YOLOv8 and DDRNet [25], respectively.

In substation environments, the background is complex and diverse, and a variety of disturbing factors may be present. Meanwhile, corrosion defects may have different scales and shapes. In order to improve the performance of the algorithm in this situation, this paper introduces multi-head attention block in YOLOv8 and DDRNet, which helps the models to distinguish between corrosion defects and background noise and improves the performance of the detection system in complex environments. Adaptively handling features of different scales makes the model performance more robust and stable when facing different types of defects. In addition, in order to solve the difficult problem of collecting samples of corrosion defects, we intercepted the corroded areas from devices with corrosion and pasted them on uncorroded devices, thus expanding the samples of corrosion defects. Through experimental validation on the self-constructed dataset, the proposed two-stage detection method yielded better real-time performance and higher accuracy than the semantic segmentation algorithm alone in locating the corroded areas in substations. Meanwhile, the improved versions of YOLOv8 and DDRNet were found to be more suitable for substation scenarios compared to the original networks. Finally, we determined that the designed data enhancement method can effectively improve the accuracy of the model for corrosion detection.

The main contributions of this paper can be summarized in the following three points:

1.

A two-stage detection method that fuses target detection and semantic segmentation is proposed to measure real-time substation corrosion regions.

2.

A multi-head attention block is used to optimize the original YOLOv8 and DDRNet networks to make them more suitable for substation scenarios.

3.

A cut-copy-paste data enhancement method is designed to expand the samples of corrosion defects.

2. Related Work

In recent years, the You Only Look Once (YOLO) family and two-branch semantic segmentation network have been widely used in many fields. This section will review the representative achievements.

2.1. You Only Look Once

The YOLO model has high detection speed and can detect objects in real-time video streams. This is because unlike traditional object detection methods such as the R-CNN family, YOLO uses an end-to-end training approach that treats object detection as a regression problem, mapping directly from image pixels to bounding boxes and class probabilities, thus simplifying the training and reasoning process. This makes YOLO particularly suitable for application scenarios that require fast response. In addition, compared to some other advanced target detection methods, YOLO’s models are typically smaller and can be more easily deployed on resource-constrained devices, such as mobile devices and embedded systems.

Due to the powerful performance of YOLO, it is used in many fields. The work in [26] presented an improved YOLOv4-Tiny algorithm for surface crack detection. The work in [27] integrated YOLOv5 into inflated 3D models that can be used for human activity recognition. In [26], a lightweight infrared object detection model G-YOLO-based on UAV was proposed. YOLO has also emerged in power inspection work, such as transmission line corridor external damage risk detection [28], insulator defect detection [29,30,31], and wind turbine blade damage detection [32]. In particular, YOLOv8, as a relatively new framework of the YOLO series, is relatively more efficient in terms of computing resources and memory consumption, so a large number of studies have begun to try to apply it to real-world scenarios. The work in [33] utilized the improved YOLOv8 algorithm in conjunction with an infrared image slicing-assisted hyper-inference technique to automatically detect Inter-Turn Short Circuit fault traces in dry-type transformers. The work in [34] presented an improved YOLOv8-based object detection algorithm for infrared images of substation equipment. The work in [35] introduced a road indicator detection model using the advanced features of YOLOv8. The work in [36] proposed an efficient small-target YOLO detection model based on YOLOv8 for Unmanned Aerial Vehicle object detection.

2.2. Two-Branch Semantic Segmentation Network

The two-branch semantic segmentation network is a deep learning model architecture that shows significant advantages in semantic segmentation tasks. A two-branch network usually contains a low-level feature branch and a high-level feature branch, where the high-level feature branch usually contains rich contextual information, while the low-level feature branch contains more detailed local information. The model uses a two-branch structure, which is able to find a balance between details and global context. Moreover, the design can effectively integrate multi-scale features so as to improve the recognition ability of the model to different scale targets.

Because of its good real-time performance, the two-branch network has been applied in the fields of automatic driving [37,38], remote sensing image analysis [39,40], and medical impact analysis [41,42]. Specifically, in autonomous vehicles, the two-branch network is used to identify roads, pedestrians, vehicles, traffic signs, etc., to help the vehicle understand the surrounding environment and make safe driving decisions. The application of two-branch semantic segmentation network in remote sensing image analysis can significantly improve the semantic segmentation accuracy of high-resolution images. Through the fusion of multi-scale features, the network can better cope with complex scenes and diversified task requirements. In the medical field, the two-branch network is used to analyze images to help doctors quickly and accurately identify and segment lesions, organs, etc., to assist diagnosis and treatment. DDRNet is a relatively new two-branch semantic segmentation network, which has been adapted by many researchers to better suit their specific tasks. In [43] DDRNet was used for real-time segmentation in unstructured environments to enable autonomous navigation for off-road robots. The use of DDRNet in [44] was used for automated driving assistance to cope with the complexity of road scenarios and the real-time requirements of segmentation algorithms in application scenarios. The work of [45,46] used DDRNet to sense the environment for the automatic unloading of outdoor forklifts.

Through a review of the above work, we can see that YOLOv8 and DDRNet stand out in many fields with their high-precision computation, compact network model, and excellent high-resolution image processing capabilities. In the particular scenario of substation inspection, in addition to high precision, the real-time performance of the algorithm is also very high. At the same time, in order to ensure refined inspection, an on-board camera and monitoring probe equipped with the substation inspection robot offer relatively high resolution.

Because YOLOv8 and DDRNet can effectively address these challenges, we chose them as the benchmark models for this work.

3. Materials and Methods

In this section, we first describe the dataset in the experiment and the corresponding data enhancement method. Then, the proposed two-stage substation corrosion defect detection method is presented, where each stage is described in a subsection. Finally, the multi-head attention mechanism block used to improve the semantic segmentation algorithm is presented.

3.1. Dataset and Data Enhancement Strategies

The image dataset used in this paper is composed of 7215 corrosion images taken by substation staff, intelligent inspection robots, and light-load head. The maximum resolution of the image is 4032 × 3024, the minimum is 2048 × 1024, and the image data were divided into a training set (5050 images), a verification set (722 images), and a test set (1443 images) in a ratio of 7:1:2. Labelme was used to mark the equipment to be tested in each picture in the form of a rectangle. In this work, there were five types of equipment that we tested, namely, cooling fin (CF), oil storage tank (OST), disconnecting switch (ds), circuit breaker (CB), and flange. Then, based on the boundary box of each marked object in the annotation file, it was cut out from the picture, and then labelme was used to carry out pixel-level boundary annotation. The semantic segmentation employed one more category, i.e., corrosion and then object detection.

We counted the number of instances of each type of marked object and the proportion of pixels occupied by each type of object. The statistical results are shown in Figure 1. As can be seen from Figure 1b, the corrosion area accounts for 0.9% of the whole dataset, accounting for the smallest proportion and 29.9% lower than the OST, having the highest proportion. Serious data imbalance will inevitably bring difficulty to the training of the model. To solve this problem, we proposed to use the cut-copy-paste method to expand the corroded sample. That is, a corrosion area is cut from the corroded equipment and then pasted onto the non-corroded equipment. In the process of capturing, the input original image and the corresponding label are captured at the same time, which can save the time of relabeling. It should be noted that only similar devices can cut and paste each other, which is done to maintain the fidelity of the corroded area. The sample augmented by this method is shown in Figure 2.

3.2. YOLOv8 Model Structure

In the first stage, we adopted the object detection algorithm YOLOv8. As the latest one-stage object detection algorithm, YOLOv8 has the advantages of high real-time performance, multi-scale detection, and global context information utilization, which is suitable for substation equipment corrosion defect detection task. YOLOv8 uses a lightweight backbone network and a series of optimization strategies to achieve fast and accurate target detection. The architecture of YOLOv8 can be divided into three key components: backbone, neck, and head, as shown in Figure 3. In the backbone stage, YOLOv8 integrates depthwise separable convolutions to improve computational efficiency and feature extraction capabilities. In the neck part, path aggregation network structure is used to improve feature fusion and capture spatial context information, which help the model better detect targets at different scales. In terms of the head, the anchor-free method is introduced to directly predict the bounding box of the target without relying on a predefined anchor, and the decoupling head structure is adopted so that the classification task and the regression task can be processed independently. The network structure of YOLOv8 is shown in Figure 3.

In this paper, in order to further improve the detection performance of YOLOv8 in the substation environment, we added a multi-head attention block (MHAB) in front of the YOLOv8 decoupling head. This improvement aims to provide a finer focus on different features, allowing the model to better capture complex features and details in the substation environment. This mechanism helps to enhance the adaptability of the model in the face of multiple substation equipment and their different states, and it ultimately improves the accuracy and robustness of the detection. The addition of MHAB to YOLOv8 is shown in Figure 3.

3.3. DDRNet Model Structure

The semantic segmentation network used in this work is DDRNet, which is composed of one high-resolution branch and one low-resolution branch, and it aims to extract high-resolution global feature information and high-level semantic details simultaneously. Suppose the output feature graph of a high-resolution branch is $F_p$, which can be expressed as

$$F_p=f_p(I;{\theta }_p),$$

(1)

where $f_p$ is the network function of the main branch, ${\theta }_p$ is the parameter of the high-resolution branch, $I\in {\mathbb{R}}^{H\times W\times C}$ is the input image, H is the image height, W is the image width, and C is the number of channels of the image. Suppose that the output feature chart of a low-resolution branch is $F_a$, which can be expressed as

$$F_a=f_a(I;{\theta }_a),$$

(2)

where $f_a$ is the network function of the main branch, and ${\theta }_a$ is the parameter of the high-resolution branch. The feature graph of the two branches is fused directly using addition. The fused feature graph $F_f$ can be expressed as

$$F_f=Add(F_p,F_a)$$

(3)

Finally, the fused feature map is mapped to the category label of each pixel through a segmentation head. Assuming that the output segmentation result is S, it can be expressed as

$$S=h(F_f;{\theta }_s)$$

(4)

where h is a function of segmentation head, and ${\theta }_s$ is parameter of segmentation head.

The design of two-branch semantic segmentation network makes it have both strong performance and excellent real-time performance. Therefore, this work used DDRNet as the benchmark network. However, the surveillance video and image of substation contain a lot of information, involving equipment, environment, personnel, and other aspects, so background noise and interference may affect the detection effect. To solve this problem, we added a multi-head attention block (MHAB) to DDRNet to form our improved version of DDRNet. An overview of the improved DDRNet and its parameter details are shown in Figure 4 and

Table 1. Details the architecture of the improved DDRNet. The improved version of DDRNet in this table assumes an image size of 896 × 896 × 3 as input. N in the table indicates the number of types of objects to be detected.

Stage	Output	DDRNet
1	448 × 448	3 × 3, 32, stride 2
	224 × 224	3 × 3, 32, stride 2
	224 × 224	$\left[\begin{array}{c}3\times 3,32\\ 1\times 1,32\end{array}\right]\times 2$
	112 × 112	$\left[\begin{array}{c}3\times 3,64\\ 1\times 1,64\end{array}\right]\times 2$
2	112 × 112, 56 × 56	$\left[\begin{array}{c}3\times 3,64\\ 1\times 1,64\end{array}\right]\times 2$	$\left[\begin{array}{c}3\times 3,128\\ 1\times 1,128\end{array}\right]\times 2$
		MHAB	–
3	112 × 112, 28 × 28	$\left[\begin{array}{c}3\times 3,64\\ 1\times 1,64\end{array}\right]\times 2$	$\left[\begin{array}{c}3\times 3,256\\ 1\times 1,256\end{array}\right]\times 2$
		MHAB	–
4	112 × 112, 14 × 14	$\left[\begin{array}{c}1\times 1,64\\ 3\times 3,64\\ 1\times 1,128\end{array}\right]\times 2$	$\left[\begin{array}{c}1\times 1,256\\ 3\times 3,256\\ 1\times 1,512\end{array}\right]\times 2$
		MHAB	SPP
5	112 × 112	Addition
	112 × 112	$\left[\begin{array}{c}3\times 3,64\\ 1\times 1,N\end{array}\right]$
	896 × 896	Bilinear Upsampling

.

The original DDRNet is mainly composed of resdidual blocks and resdidual bottleneck blocks. We only inserted the MHAB in the high-resolution branch, because the feature map in the high-resolution branch contains more information and is more conducive to extracting the contour information of the target. The precise location of the MHAB in the improved version of DDRNet is shown in Figure 4. In addition, the space pyramid pool (SPP) uses DDRNet’s original deep aggregation pyramid pool module.

3.4. Multi-Head Attention Block

The structure of the Muti-Head Attention Block is shown in Figure 5. The central idea of the multi-head attention mechanism is that multiple “attention heads” process different parts of information in parallel to capture different features and relationships. Each “head” can be thought of as an independent attention mechanism that learns and focuses on different aspects of the input data in different subspaces. First, for the input, we map it to different Query, Key, and Value vectors through linear transformations. For the ith attention head, we have the query vector $Q_i=X{W}_i^Q$, the key vector $K_i=X{W}_i^K$, and the value vector $V_i=X{W}_i^V$, where $W_i^Q$, $W_i^K$, and $W_i^V$ define the weight matrix corresponding to each attention head. For each attention head i, we calculate the attention weight. The dot product of the query vector and the key vector are first calculated; they are then scaled and converted to a probability distribution by the softmax function. The formula is

$$Attention\_Score{s}_i={\displaystyle \frac{Q_i{K}_i^T}{\sqrt{d_k}}},$$

(5)

where $d_k$ is the dimension of the key vector (usually $d_k=d_{model}/h$, where h is the number of attention heads). Then, we apply the softmax function:

$$Attention\_Weight{s}_i=softmax(Attention\_Score{s}_i)$$

(6)

Next, we use the calculated attention weight to weight the value vector:

$$Outpu{t}_i=Attention\_Weight{s}_i{V}_i$$

(7)

4. Experiments and Analyses

4.1. Implementation Details

The operating system used for the experiments was Windows 10, and Pytorch 2.1 and GPU acceleration were used for training and testing. The GPU model was NVIDIA GeForce RTX 4060 (8 G); the software environments were Anaconda3, Cuda12.2, and Python 3.9; and the development tool was PyCharm Community Edition 2022.2.2.

In the proposed method, the training processes of the target detection and semantic segmentation were independent of each other, and the loss functions involved in the training process of both are described below. In the training process of target detection, the loss function was divided into two types: a classification loss function and a bounding box loss function. In YOLOv8, the classification loss function uses the cross-entropy loss. The bounding box loss function uses a combination of a Distribution Focal Loss (DFL) loss function and a CIOU loss function, where the DFL calculates the offsets from the center of the anchor point to the upper-left and lower-right corners, and the CIOU calculates the localization loss of the positive samples only.

To ensure that the training can be carried out smoothly, the initial learning rate for substation equipment type identification using the YOLOv8 network in the first stage was set to 0.005, while in the second stage, the initial learning rate for equipment and corrosion defect segmentation using the DDRNet network was set to 0.0005. In both stages, the number of epochs was set to 500. The Adam function was used as the optimization function in both stages. The image size for training and testing was uniformly reduced to 1500 × 1500.

4.2. Evaluation Indicators

This section introduces the evaluation indicators used in the experiment, which are defined as follows:

Mean Average Precision ($mAP$) is a common evaluation index used in object detection tasks to measure the performance of a model in detecting different types of objects. For each category of targets, we first need to calculate the Precision and Recall of that category. The accuracy rate represents the proportion of correctly predicted positive samples to all predicted positive samples, and the recall rate is the proportion of correctly predicted positive samples to all true positive samples. By changing the confidence threshold of the detection model, different Precision rates and Recall rates can be obtained to form a precision–recall ($PR$) Curve. $AP$ is the area under the entire $PR$ curve, and the $PR$ curve is generally integrated to calculate the area. After calculating the $AP$ in all categories, the $AP$ of all categories is averaged to obtain $mAP$. The formula is

$$mAP={\displaystyle \frac{1}{N}}\sum _{i=1}^{N}A{P}_i$$

(8)

where N is the total number of categories, and $A{P}_i$ is the $AP$ value of the i-th category.

Mean pixel accuracy ($mAcc$) is the ratio between the number of pixels correctly classified by the model and the total number of pixels, which is defined as follows:

$$Acc={\displaystyle \frac{T_p+T_n}{T_p+T_n+F_p+F_n}}$$

(9)

$$mAcc=\sum Acc/\phantom{AccN}N$$

(10)

In this equation, N represents the number of categories in the semantic segmentation task, $T_p$ means true positive, $T_n$ means true negative, $F_p$ means false positive, and $F_n$ means false negative.

Mean Intersection over Union ($mIoU$) measures the degree of overlap between the predicted and true regions, while ($mIoU$) is the average of all categories of ($IoU$). For each category, the ($IoU$) is calculated as follows:

$$IoU={\displaystyle \frac{T_p}{T_p+F_p+F_n}}$$

(11)

where the meanings of $T_p$, $F_p$, and $F_n$ are the same as above. After calculating the ($IoU$) for each category, we take the average of these IoU values to obtain the ($mIoU$):

$$mIoU=\sum IoU/\phantom{IoUN}N$$

(12)

4.3. Comparison with Existing Methods

In order to verify the effectiveness and advantages of the proposed method, we compared it with existing target detection and semantic segmentation algorithms. The mainstream object detection algorithms Sparse R-CNN [47], Faster RCNN, and YOLOX [18], as well as the semantic segmentation algorithms DeepLabV3 [48], Mask2Former [49], UNet [22], and DDRNet, were selected for comparison. All algorithms were fully trained on self-built datasets. Due to space constraints, only the variation in the loss function of our methods (i.e., YOLOv8 and DDRNet with MHAB inserted) during the training process is shown here, as shown in Figure 6. The final experimental results are shown in

Table 2. Comparative experiments with existing object detection algorithms.

Method	mAP
Sparse R-CNN	53.46
YOLOX	48.97
Faster R-CNN	37.25
YOLOv8	56.53
YOLOv8 & MHAB	59.18

and

Table 3. Comparative experiments with existing semantic segmentation algorithms.

Method	mIoU	mACC
DeepLabV3	70.83	75.76
UNet	65.48	67.62
Mask2Former	73.94	74.28
DDRNet	72.24	73.95
DDRNet & MHAB	75.36	78.64

. The results show that the performance of YOLOv8 and DDRNet improved by MHAB was the best. Specifically, YOLOv8 with MHAB improvement performed 4.6% better on the mAP than the YOLOv8 without improvement, as shown in Table 2; DDRNet with MHAB improvements performed 4.3% better on mIoU and 6.3% better on mACC than DDRNet without improvements, as shown in Table 3.

4.4. Ablation Experiment

The purpose of this part was to verify the validity of the proposed data enhancement method. We first trained with the unenhanced dataset and then retrained the model with the enhanced dataset. The algorithm used was DDRNet modified by MHAB. The results of the two training sessions are shown in

Table 4. Experimental results of DDRNet modified by MHAB under data enhancement and without data enhancement.

Enhancement	Corrosion		CF		OST		DS		CB		Flange
	IoU	ACC	IoU	ACC	IoU	ACC	IoU	ACC	IoU	ACC	IoU	ACC
✗	45.65	46.37	86.29	87.16	79.85	81.67	72.45	76.32	63.63	66.58	60.24	61.29
✔	56.32	57.61	85.97	86.19	80.62	82.49	75.89	77.93	65.98	67.46	61.65	65.81

. It can be seen that although the algorithm improvement after data enhancement training was not obvious in the non-corrosive category of IoU (CF, OST, DS, CB, and flange), the performance of the network after data enhancement training was 23.37% and 24.24% higher than that of the network without data enhancement training in the Corrosion category of IoU and ACC, respectively. Therefore, the experiment shows that the cut-copy-paste enhancement strategy can significantly improve the model accuracy under unbalanced sample conditions.

4.5. Real-Time Comparison of Algorithms

The purpose of the proposed two-stage method was to reduce the time consumed by the model during inference. Therefore, this part of the experiment proved the real-time performance of this method. The final experimental results are shown in

Table 5. Comparison of calculation time of different methods.

Method	Algorithm	Time (ms)
One stage	DDRNet	11.28
Two stage	YOLOv8	1.35
	Postprocessing	0.86
	DDRNet	3.58

. It can be seen that although DDRNet had a faster inference speed, it was still not as good as the well-designed two-stage detection method, because DDRNet performs pixel-level calculation on the whole image, while the two-stage method uses a YOLOv8 to filter part of the background first.

4.6. Visual Result Discussion

In order to verify the accuracy of the experimental results more intuitively, we carried out a visual display of the experimental results. Meanwhile, in order to comprehensively compare the performance of different methods, we have presented the results of all existing methods involved in the experiment, including Sparse R-CNN, Faster R-CNN, YOLOX, YOLOv8, DeepLabV3, Mask2Former, UNet, and DDRNet. The results also include our improved YOLOv8 and DDRNet with MHAB.

The results of various object detection methods in stage 1 are shown in Figure 7. It can be seen that although the bounding boxes predicted by these methods were relatively close to the labels, our proposed YOLOv8 and MHAB had the highest confidence. In the semantic segmentation results of stage 2 (as shown in Figure 8), although the methods performed similarly in the overall segmentation of the device, DDRNet and MHAB performed better in the segmentation of finer corrosion areas. This is because the attention mechanism it introduces allows the algorithm to focus on the target area more efficiently. Through visualization comparison of the methods involved in the experiment, not only the effectiveness of the proposed method is verified, but also its superiority is further highlighted.

4.7. Failure Case Discussion

Although our method has been effective to a certain extent, it still has some shortcomings. Especially in the case of semantic segmentation in backlight, the shadow region may be misjudged as a corrosive region (as shown in Figure 9), which is undoubtedly a key point that needs to be improved. To effectively address this issue, we plan to first collect a large amount of data under various lighting conditions (especially in backlit scenes) to ensure that the model is exposed to a sufficient number of backlit and shadowed samples during training. Subsequently, we will annotate the shadows and eroded regions more finely, so as to ensure that the model can accurately distinguish the difference between the two during training.

5. Conclusions

This work has proposed a two-stage method based on YOLOv8 and DDRNet to detect the corrosion defects of substation equipment. In the first stage, the improved target detection network YOLOv8 was used to conduct object detection on the equipment in the substation. In the second stage, a DDRNet semantic segmentation network was used to accurately segment substation equipment and its corrosion defects. In addition, multi-head attention block (MHAB) was used in YOLOv8 and DDRNet, effectively improving the detection capabilities of both in substation environments. Finally, a clip-copy-paste method was used to expand the corrosion sample. Experimental results show that the proposed two-stage substation corrosion location method can meet good real-time performance, and the effectiveness of MHAB and data enhancement was also verified. However, the proposed method is prone to misidentify the shaded area as a corroded area under backlight conditions. Therefore, it is planned that the shadows and erosion areas will be labeled in future work to ensure that the model can accurately distinguish the difference between the two when trained.