Surface defect detection of ceramic disc based on improved YOLOv5s

Abstract

Addressing the challenges in detecting surface defects on ceramic disks, such as difficulty in detecting small defects, variations in defect sizes, and inaccurate defect localization, we propose an enhanced YOLOv5s algorithm. Firstly, we improve the anchor frame structure of the YOLOv5s model to enhance its generalization ability, enabling robust defect detection for objects of varying sizes. Secondly, we introduce the ECA attention mechanism to improve the model's accuracy in detecting small targets. Under identical experimental conditions, our enhanced YOLOv5s algorithm demonstrates significant improvements, with precision, F1 scores, and mAP values increasing by 3.1 %, 3 %, and 4.5 % respectively. Moreover, the accuracy in detecting crack, damage, slag, and spot defects increases by 0.2 %, 4.7 %, 5.4 %, and 1.9 % respectively. Notably, the detection speed improves from 232 frames/s to 256 frames/s. Comparative analysis with other algorithms reveals superior performance over YOLOv3 and YOLOv4 models, showcasing enhanced capability in identifying small target defects and achieving real-time detection.

1. Introduction

In the production process of daily ceramic plates, various defects occur at different stages such as molding, pressing, drying, washing, glazing, and firing [1]. The surface quality of ceramic plates affects their aesthetics and usability. The surface quality of ceramic plates significantly impacts their aesthetics and usability. Despite the high degree of automation achieved by daily ceramic production enterprises in key processes such as throwing, glazing, and firing, manual subjective judgment and inspection still dominate the quality inspection process. Relying solely on manual visual inspection methods is time-consuming and labor-intensive, posing challenges to meeting the high-efficiency requirements of ceramic manufacturing. The curved surface characteristics of ceramic discs exacerbate issues such as high light reflection and uneven illumination. Small defects may be concealed within highlight or shadow areas, and the size of the defective area varies, presenting a significant challenge for daily ceramic disc defect detection.

At present, numerous object target detection and surface defect detection methods rely on traditional digital image processing techniques. These techniques have found wide applications in diverse domains, including remote sensing image recognition, crystal oscillator defect detection, and inspection of weld seams and internal defects. Within the research field of surface defect detection in daily ceramics, various methods have been proposed. For instance, Xie et al. introduced a machine vision-based surface defect detection method [2].

This method initially applies image morphological closing operations to acquire the regions of interest (ROIs) within the image. Subsequently, the ROI region undergoes a comparison with the ROI processed by the separated median filtering algorithm to identify spot defect information on the ceramic plate surface. However, this algorithm exhibits limited generality and is capable only of detecting spot defects. Feng et al. proposed a detection method based on the principle of stereo matching algorithm, combined with slag characteristics, to detect slag on the surface of ceramic plates using the disparity map as the segmentation basis [3].This method can effectively identify slag on the surface of ceramic plates but has not been validated for identifying other defects. Lin et al. proposed a ceramic plate defect detection algorithm based on multiscale feature fusion [4]. The method calculates image features using multi-scale gray histogram and the Sobel operator, and then utilizes a KNN classifier for training after fusion to obtain the classification model. Although this method effectively identifies three types of defects, namely spots, cracks, and damage, with an average recognition accuracy of 86.86 %, it does not account for slag defects, and the average recognition accuracy requires improvement. Although the aforementioned algorithms address certain surface defect detection challenges in ceramics, they are limited in their ability to detect various defect types, are sensitive to abnormal lighting conditions, and lack applicability and generality. Additionally, traditional image processing algorithms depend on manually designed feature extraction operators, which are cumbersome and complex. Moreover, they demonstrate insufficient robustness and low detection rates in industrial application scenarios.

Currently, the majority of research on ceramic tile defect detection focuses on deep learning-based visual detection methods, while there is comparatively less research on daily ceramic defect detection. The balanced multiscale target scoring network algorithm, proposed by Cao et al. for ceramic tile surface defect detection, enhances the YOLOv5s algorithm by introducing content-aware feature recombination and dynamic attention mechanisms [5].The proposed method achieved an average precision improvement of 4.9, with a 6 % increase in AP for small targets. A supervised automatic detection method for ceramic tile surface defects, proposed by Lu et al., introduces depth separable convolution, incorporates DWBottleneck and residual connections, and replaces the feature extraction backbone of YOLOv5s with an improved Shufflenetv2 [6]. Additionally, it enhances the original YOLOv5s algorithm by integrating the SE channel attention mechanism. Compared to the original YOLOv5s detection algorithm, the proposed method reduces model parameters by 20.46 % and floating-point operations by 26.22 %. The mean average precision (mAP) is 96.73 %, representing a 1.93 % improvement over the original YOLOv5s. Wan et al. proposed ceramic tile surface defect detection based on deep learning [7]. A lightweight tile detection system is constructed by adding an attention mechanism CBAM module to the YOLOv5s network structure, adding a small-scale detection layer, and replacing the convolution with a depthwise separable convolution. Compared with the original YOLOv5s model, the accuracy, F1 and mAP of the proposed method are increased by 6.09 %, 6.53%and 6.15,% respectively.

The common feature between tiles and everyday ceramic products lies in their surface glossiness, which enables them to reflect light. However, due to the curved surface characteristics of everyday ceramics, there are variations in glossiness across the surface. The uneven distribution of light on the surface of everyday ceramics often conceals small defects, and areas with high glossiness may be mistakenly identified as defects. Therefore, deep learning models based on tile detection are not suitable for detecting surface defects on everyday ceramics.

Addressing the limitations of existing methods, we focus on ceramic plates and utilize YOLOv5s as the foundational deep learning model. We aim to detect and recognize four types of defects on the surface of ceramic plates: cracks, breakage, residue, and spots. To mitigate the interference of high-gloss areas on defect features and enhance the model's detection generalization, we treat high-gloss areas as detection targets as well. The contributions of this study can be summarized as follows: Our team actively collaborated with Jingdezhen Hongye Ceramic Factory to establish a surface defect detection platform for ceramic plates on the production line. We manually annotated a total of 2404 sample labels, including 5 detection targets, with small targets accounting for 52.26 % of the annotated dataset. Our research is based on the YOLOv5s model, and we made improvements to the network structure. Firstly, we improve the anchor frame structure of YOLOv5s, and design a new anchor frame to replace the original one according to the data characteristics of our labeled training samples. Then, the ECA attention mechanism module is added to the backbone network of YOLOv5s model. The experimental results verify the feasibility of the proposed method, which provides a reference for the surface defect detection task of daily ceramic products.

2. YOLOv5 model introduction and network structure composition

YOLOv5 is an object detection model proposed by the Ultralytics team, which is based on YOLOv4 with some new improvements. YOLOv5 provides four network architectures, from small to large, allowing a trade-off choice between speed and accuracy. YOLOv5s has the smallest network structure, which is more prominent for small object detection and more suitable for small object detection. Therefore, the network structure of YOLOv5s is improved in this study, and the network structure of YOLOv5s is shown in Fig. 1 [8].

Fig. 1.

Network structure diagram of YOLOv5s.

In the Input side, Mosaic data enhancement strategy stitches four images together in different combinations to increase the diversity of training data and improve the diversity of the model in the face of complex shape conditions [9]. Adaptive anchor boxes learn dynamically through the training process to better adapt to targets of different sizes and aspect ratios. The adaptive image scaling method is used to unify the input size of the network, so that the amount of calculation is reduced in the inference process of the model.

The backbone network of YOLOv5s uses Focus structure and CSP structure. The main purpose of the Focus structure is to reduce the amount of computation by decomposing the larger feature map into four smaller feature maps while retaining important information, as shown in Fig. 2 [10]. The CSP structure splits the feature map into two parts, applies a convolution operation on one part, and joins the other part with the result of the previous convolution operation [11].

Fig. 2.

Slicing operation.

The Neck end of YOLOv5s adopts FPN and PAN structures for feature fusion and dimension reduction, which helps to improve the detection ability of the network for small objects [12].

The Prediction part in YOLOv5s is composed of a series of convolutional layers and a final output layer. The convolution structure is responsible for extracting the features of different levels, and the output layer maps the features to grid cells of different scales, so as to realize the detection of objects of different sizes [13].

3. 3. the improved YOLOv5s model

3.1. Network improvement with ECA module of attention mechanism

3.1.1. An introduction to ECA attention mechanisms

When observing a global image, humans tend to focus more attention on the area of interest to suppress other irrelevant information, thus obtaining more specific information about the observed object more quickly. Attention mechanisms in target detection networks are essentially the same as human selective visual attention mechanisms [14]. Introducing attention mechanism into the target detection network can strengthen the representation ability of the feature map in the network and improve the detection accuracy of the network [15].

Since the defect targets in the ceramic disc are usually small, in order to improve the recognition accuracy of the model, the Efficient Channel Attention (ECA) module is added to the feature extraction module in the YOLOv5s network structure, which can detect the area where the small target defects are located and suppress the irrelevant background region.

3.1.2. The improved backbone network structure

The backbone network of YOLOv5s, which consists of Focus, CBL, BottleneckCSP, and SPPF modules, is shown in Fig. 3.

Fig. 3.

YOLOv5s backbone network structure.

To optimize the network, the ECA attention mechanism module is embedded with CSP1_1, CSP1_2, and CSP1_3 module to replace the original BottleneckCSP module. The diagram of improved network structure is shown in Fig. 4.

Fig. 4.

The improved backbone network structure diagram.

In this paper, for the problem of small target detection of ceramic discs, the ECA attention mechanism module is introduced into YOLOv5s, which brings two benefits, the first one is to enhance the network's ability to extract critical and important information in the face of defects detection such as cracks, damages, slag, and spots, so that the network can make a more accurate identification, and the second one is to effectively reduce the overhead brought by the model in the computation and storage, and to accelerate the detection speed.

3.1.3. ECA attention mechanism principle

The ECA attention mechanism works as shown in Fig. 5, where ECA first performs a global average pooling operation on each channel to reduce the feature values within each channel to a scalar. Next, ECA computes an adaptive convolution kernel size based on the scalar values on each channel. The 1-D convolution operation is then performed on each channel using the adaptively calculated convolution kernel size, resulting in a weight for each channel. The weight obtained by the convolution operation is applied to each channel of the feature map, so that the feature information between different channels is properly concerned and adjusted [16].

Fig. 5.

ECA network structure.

ECANet uses one-dimensional convolution to realize cross-channel information interaction. The size of the convolution kernel adaptively changes according to the specific function, so that layers with more channels can interact more fully across channels [17]. The adaptive function is shown in equation (1). Where γ = 2 and b = 1.

$$k=\left|\frac{{\log }_2\left(c\right)}{\gamma }+\frac{b}{\gamma }\right|$$ (1)

In this paper, the original C3 module is changed into an ECA module, the Model Summary of YOLOv5s network is reduced from 270 layers to 225 layers, and the weight model size is reduced, the number of network parameters is reduced, and the speed of model training and inference is accelerated.

3.2. Anchor frame optimization

3.2.1. Anchor frame definition

Anchor boxes refer to a set of predefined boxes defined on an image, used for searching target objects at different positions and scales [18]. In YOLOv5, the selection of anchor boxes is typically adjusted based on the characteristics of the dataset to better adapt to the sizes of target objects. In the original YOLOv5, when the input image size is 640x640, three different-sized feature maps are obtained, namely 80*80(640/8), 40*40(640/16), and 20*20(640/32) [19]. The anchor box matrix is defined as shown in Equation (2).

$$\begin{array}{c}\left[10,13,16,30,33,23\right]\\ \left[30,61,62,45,59,119\right]\\ \left[116,90,156,198,373,326\right]\end{array}$$ (2)

The first row of the anchor box matrix corresponds to the 80*80 feature map, the second row corresponds to the 40*40 feature map, and the third row corresponds to the 20*20 feature map [20]. Each row of the anchor box matrix contains three pairs of coordinates, representing the coordinates of the top-left and bottom-right corners of the anchor box. These anchor boxes are typically generated at different positions on different feature maps to cover objects of various scales. In YOLOv5, these anchor boxes are associated with each grid cell on the feature map and used to predict whether there is an object within that grid cell, as well as the position and class information of the object [21]. Specifically, for each grid cell, the model predicts whether one of the multiple anchor boxes contains an object and makes predictions regarding the object's position and class.

The anchor frame structure of the original YOLOv5 cannot be adapted to other object target detection. In our training sample data distribution, the defects of a considerable part of samples show a long-tail distribution. We counted the distribution of the aspect ratio of all target calibration frames in the data set and drew a histogram, as shown in Fig. 6. However, we noticed that there were 2915 targets in the training samples whose aspect ratio was between 0 and 2. For example, when labeling defects such as slag falling, the labeling box ratio was 1, while the labeling box ratio of defects such as cracks was 2 or larger. The ratio of labeled boxes is often not the same for different targets, as shown in Fig. 7.

Fig. 6.

Diagram of the corresponding labeled boxes for different targets.

Fig. 7.

Marking box aspect ratio histogram.

To this end, we improve the YOLOv5 anchor structure and set a new anchor structure to adapt to sample training.

3.2.2. Anchor frame improvement method

The specific method is to obtain the annotation information of the image from the data set, which includes the position and size of various objects in the image, and calculate the width and height range of the anchor box according to the annotation information [22]. The labeled boxes are then clustered using the K-means clustering algorithm. In this process, the algorithm will automatically determine the best cluster center and generate a specified number of anchor boxes according to the size and distribution of the annotated boxes [23]. The size and shape of these anchor boxes can well represent the size distribution of the target objects in the dataset. The genetic algorithm is used to optimize the generated anchor box to further improve the quality of the anchor box. Finally, the optimized anchor box is returned for YOLOv5 model training and inference. These anchor boxes are able to cover objects of different sizes and shapes, and can maximize the detection performance. The improved anchor box matrix is defined as shown in Equation (3).

$$\begin{array}{c}\left[38,36,53,49,47,92\right]\\ \left[71,64,121,78,87,114\right]\\ \left[206,106,131,176,251,243\right]\end{array}$$ (3)

4. Experiment and analysis

4.1. Experimental environment and training parameters

The experiment environment collocated in this paper is Windows 10 operating system, and the detailed experiment configuration is shown in Table 1. The model training parameters are shown in Table 2.

Table 1.

Experiment configuration environment.

Parameter names	Configuration parameters
Operating system	Windows10
Programming languages	Python
Independent graphics card	RTX4070
Deep Learning Framework	Pytorch
Cuda version	Cuda 12.2.79

Table 2.

Model training parameters.

Parameter names	Configuration parameters
Input size	640*640
batch size	16
epochs	300
momentum	0.937
Mosaic lr0 warmup epochs warmup momentum	1 0.001 20 0.8

4.2. Dataset description

The data set is collected on the production line of Jingdezhen Hongye Ceramics Factory. As shown in Fig. 8, five categories of target images are collected, including spot, highlight, crack, damage and slag. A total of 2404 sub-photos are manually labeled, and 85 % of the data set is divided into training set, 10 % as validation set, and the remaining 5 % as test set. Table 3 shows the target distribution of each class of data.

Fig. 8.

Acquisition target diagram.

Table 3.

Distribution of various types of targets.

Types	Training set	Validation set	Test set	Total
Total number of images	2043	120	241	2404
Proportion of the number of images	85 %	10 %	5 %	100 %
Number of crack targets	511	37	65	613
Number of broken targets	571	39	58	668
Target quantity of slag	641	29	89	759
Number of spot targets	551	37	61	649
Number of highlight targets	619	37	78	734
Total number of dataset targets	2893	179	351	3423

According to the definition of small target size for MSCOCO dataset proposed by Microsoft, when the target area is smaller than 32*32 pixels in size it is considered as a small target, target area between 32*32 pixels and 96*96 pixels is a medium-sized target, and target area larger than 96*96 pixels in size is a large-sized target. As shown in Fig. 9, the percentage of small targets in this dataset is 52.26 %, the percentage of medium targets is 44.50 %, and the percentage of large targets is 3.24 %.

Fig. 9.

Distribution of target proportion.

4.3. Evaluation metrics

In order to verify whether the model can effectively detect ceramic disk defects and evaluate the model performance, we use the evaluation metrics P, AP, mAP, F1, model size, FPS. The index calculation method is shown in Equations (4), (5), (6), 7and 8. P is the number of true examples, FP is the number of false positive examples, FN is the number of false negative examples, and N is the number of categories. P, R, AP, F1, mAP and FPS represent precision, recall, PR curve area, harmonic mean of P and R, AP mean, and frames per second, respectively.

$$P=\frac{TP}{TP+FP}\times 100\%$$ (4)

$$R=\frac{TP}{TP+FN}\times 100\%$$ (5)

$$\text{AP}={\int }_0^{1}P\left(R\right)\text{dR}\times 100\%$$ (6)

$$mAP=\frac{\sum AP}{N}$$ (7)

$$F1=\frac{2TP}{2TP+FP+FN}\times 100\%$$ (8)

4.4. Comparison with other object detection models

In order to verify the superiority of the proposed algorithm, it is compared with YOLOv4 and YOLOv3 target detection algorithms. The experimental results are shown in Table 4.

Table 4.

Experimental results of different models.

Model	P(%)	F1(%)	mAP(%)	Model size(MB)	FPS
YOLOv5s + ECA + anchor	93.4	91	95.1	12.7	256
YOLOv5s	90.3	88	90.6	13.8	232
YOLOv4	85.5	62	75.2	244	168
YOLOv3	89.6	86	89.2	16.6	129

By comparing the experimental results in Table s4 and it is found that the performance of our improved YOLOV5s algorithm is better than that of other algorithms. Compared with YOLOV3, the precision rate is increased by 3.8 %, F1 is increased by 5 %, mAP value is increased by 5.9 %, and the detection speed is increased by 127FPS. Compared with YOLOV4, the accuracy is increased by 7.9 %, F1 is increased by 29 %, mAP value is increased by 19.9 %, and the detection speed is increased by 88FPS.

4.5. Ablation experiments

In order to further verify the importance of each module, we set up three groups of control experiments: the first group is the original YOLOv5s model, the second group is the YOLOv5s model with improved anchor structure, and the third group is the model with ECA attention mechanism added on the basis of improved anchor structure. In the three experiments, the dataset and training parameters were the same. The experimental results of each model are shown in Table 5. Each model corresponds to various target P and AP values as shown in Table 6.

Table 5.

Experimental results of ablation experiments.

Model	Original	ANCHOR	ECA	P(%)	F1(%)	mAP(%)	Model size(MB)
YOLOv5s	✓			90.3	88	90.6	13.8
YOLOv5+ANCHOR	✓	✓		92.7	91	92.7	12.7
YOLOv5+ANCHOR + ECA	✓	✓	✓	93.4	91	95.1	12.7

Table 6.

Experimental results of ablation experiments.

Target type	YOLOv5s		YOLOv5+ANCHOR		YOLOv5+ANCHOP + ECA
	P(%)	AP(%)	P(%)	AP(%)	P(%)	AP(%)
crack	95	96.4	95	96.6	95.2	96.8
damaged	86.5	85.1	90.3	88.4	91.2	96
slag	84.7	85.4	89.7	91	90.1	93.3
spot	92.8	93.2	95.2	94.1	94.7	93.5
highlight	92.5	92.8	93.4	93.3	95.7	96.1

4.5.1. Results and analysis of YOLOv5s + ANCHOR

In this experiment, the anchor frame structure is improved based on the YOLOv5s model. Recognition results Compared with the original model, the crack precision remains unchanged, the AP value increases by 0.2 %, the damage precision increases by 3.8 %, the AP value increases by 3.3 %, the slag precision increases by 5 %, the AP value increases by 5.6 %, the spot precision increases by 2.4 %, the AP value increases by 0.9 %, and the highlight precision increases by 0.9 %. The AP value is increased by 0.5 %. In the total recognition results, the precision, F1 and mAP are increased by 2.4 %, 3 % and 2.1 %, respectively.

The rationale behind our refinement of the anchor box structure in YOLOv5s essentially lies in adjusting the model's perceptual capability towards objects, bringing it closer to the distribution of real-world data. This enhancement notably boosts the model's performance in tackling object detection tasks of varied sizes and shapes, as evidenced by improvements in accuracy and Average Precision (AP) across all object categories.

The optimized anchor box structure enables the model to more accurately capture objects of diverse sizes and shapes, thereby enhancing precision and AP values for these categories. The improved anchor box structure achieves this by better covering the range of sizes and shapes of various objects, facilitating the model's identification of them.

Furthermore, the refined anchor box structure leads to an overall enhancement in precision, F1 score, and mAP. This indicates superior performance of the model across all object detection tasks, not limited to specific categories. The reason lies in the enhanced universality and generalization of the model towards objects, facilitated by the improved anchor box structure, allowing it to better adapt to diverse scenes and data distributions.

4.5.2. Results and analysis of YOLOv5s + ANCHOR + ECA

In this experiment, the ECA attention mechanism module is introduced based on the YOLOv5s + ANCHOR model. Compared with the YOLOv5s + ANCHOR model, the precision of crack, damage, slag and highlight is increased by 0.2 %, 0.9 %, 0.4 % and 2.3 %, respectively, and the AP value is increased by 0.2 %, 7.6 %, 2.3 % and 2.8 %, respectively. Compared with the YOLOv5s + ANCHOR model, the precision of spots is reduced by 0.5 %, but compared with the original model, it is increased by 1.9 %. Overall, the precision and mAP are increased by 0.7 % and 2.4 % respectively.

The reasons are as follows: Various defects or targets exist on the surface of ceramic plates, exhibiting different features at different positions and scales within images. The ECA attention mechanism assists the model in identifying and focusing on the crucial features relevant to the task, thereby enhancing the accuracy of target detection. In ceramic plate surface target detection, the spatial positions of features may vary, with defects or targets appearing at different locations or orientations within images. By operating on the channel dimension without altering the spatial distribution of feature maps, the ECA attention mechanism maintains spatial location invariance of features. This implies that even if defects or targets change their positions within images, the model can still effectively attend to these features, thereby improving model robustness. Given that ceramic plate surface target detection typically involves processing large amounts of image data, computational efficiency of the model is crucial. The ECA attention mechanism achieves higher computational efficiency by simply weighting each channel of feature maps without requiring additional parameters or complex computational operations. This enables the model to process image data faster while maintaining performance, thereby enhancing detection speed and real-time capabilities.

The comparison of PR curves between YOLOv5s + ANCHOR + ECA and YOLOv5s models is shown in Fig. 10. Combined with Tables 5 and in general, compared with the original model, the improved model improves the precision by 3.1 %, F1 by 3 %, and mAP value by 4.5 %. By comparing the three groups of ablation experiments, it can be verified that our improved model is superior to the original model in detection accuracy and detection speed.

Fig. 10.

Comparison of model PR curves before and after improvement.

4.6. Display of recognition results

The trained model was used to detect 241 ceramic disc pictures in the test set, and the detection results are shown in Fig. 11, where spot defects are detected in (a) figure, crack defects in (b) figure, breakage defects in (c) figure, highlight areas in (d) figure, and slag defects in (e) figure. It can achieve ideal detection effect for small target defects on ceramic disc surface.

Fig. 11.

Detection effect (a) Spot detection results (b) Crack detection results (c) Damage detection results (d) Highlight detection results (e) Slag dropping test results.

5. Conclusion

Aiming at the problem of small target defect detection of ceramic discs, a YOLOv5s algorithm with improved anchor frame structure and ECA attention mechanism module is proposed. Firstly, the anchor structure of YOLOv5s is improved, and the original anchor structure is replaced by a new anchor frame according to the data characteristics of the training samples, which enhances the robustness of the model and can effectively detect objects of different sizes. Then, the ECA attention mechanism module is introduced, and the ECA module is fused with the CSP1_1, CSP1_2, and CSP1_3 modules respectively to replace the BottleneckCSP module of the original network. The network is guided to focus on small target areas to better learn defect features, and the regression accuracy and robustness of the model are further improved. Through the collected ceramic disc defect data set, the improved YOLOv5s algorithm after training is used to identify the ceramic disc defects. The final precision, F1 and mAP are 93.4 %, 91 % and 95.1 %, respectively, and the detection speed is as high as 256 frames/S. After experimental analysis, it can be verified that the algorithm is superior to the original YOLOv5 algorithm in detection speed and accuracy, and the precision, F1 and mAP values are higher than those of YOLOv4 and YOLOv3 models. The improved algorithm has good detection performance and can be applied to ceramic automatic production line, improve production efficiency, and has good promotion value. While the improved algorithm proposed in this study has achieved significant performance enhancements in surface defect detection of ceramic discs, its applicability may be subject to certain limitations. For instance, the effectiveness of its application on other types of ceramic products may vary and would require validation and adjustment based on specific circumstances.

Data availability statement

All raw data used in this study have been backed up with appropriate storage mechanisms and are available to researchers. Data set has been uploaded to the AI Studio, can through the following links for: https://aistudio.baidu.com/datasetdetail/277762.

CRediT authorship contribution statement

Haipeng Pan: Writing – review & editing, Project administration. Gang Li: Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Hao Feng: Writing – review & editing, Project administration. Qianghua Li: Methodology, Investigation. Peng Sun: Methodology, Investigation. Shujia Ye: Methodology, Investigation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.