DVCW-YOLO for Printed Circuit Board Surface Defect Detection

Abstract

The accurate and efficient detection of printed circuit board (PCB) surface defects is crucial to the electronic information manufacturing industry. However, current approaches to PCB defect detection face challenges, including large model sizes and difficulties in balancing detection accuracy with speed. To address these challenges, this paper proposes a novel PCB surface defect detection algorithm, named DVCW-YOLO. First, all standard convolutions in the backbone and neck networks of YOLOv8n are replaced with lightweight DWConv convolutions. In addition, a self-designed C2fCBAM module is introduced to the backbone network for extracting features. Next, within the neck structure, the C2f module is substituted with the more lightweight VOVGSCSP module, thereby reducing model redundancy, simplifying model complexity, and enhancing detection speed. By enhancing prominent features and suppressing less important ones, this modification allows the model to better focus on key regions, thereby improving feature representation capabilities. Finally, the WIoU loss function is implemented to replace the traditional CIoU function in YOLOv8n. This adjustment addresses issues related to low generalization and poor detection performance for small objects or complex backgrounds, while also mitigating the impact of low-quality or extreme samples on model accuracy. Experimental results demonstrate that the DVCW-YOLO model achieves a mean average precision (mAP) of 99.3% and a detection speed of 43.3 frames per second (FPS), which represent improvements of 4% and 4.08%, respectively, over the YOLOv8n model. These results confirm that the proposed model meets the real-time PCB defect detection requirements of small and medium-sized enterprises.

1. Introduction

Printed circuit boards (PCBs) are extensively employed in integrated circuits. However, the inherent instability of the production process, coupled with environmental factors and equipment conditions, frequently leads to the emergence of various defects, including missing holes, rat bites, and open circuits. These defects have a detrimental impact on both product performance and safety [1]. Consequently, there is a growing demand for efficient, accurate, and cost-effective PCB defect inspection systems.

In recent years, deep learning-based models for PCB surface defect detection have garnered significant attention. Current PCB defect detection methods using vision technology can be broadly classified into two categories: two-stage object detection algorithms and single-stage object detection algorithms [2]. Two-stage object detection algorithms first locate the target and generate detection boxes, followed by the classification of the detected boxes to refine the target categories and positions. Notable models include the region-based convolutional neural networks (R-CNNs) [3,4,5]. For example, Ren et al. [5] developed an object detection framework that combines the strengths of R-CNN and a new region proposal network (RPN) to improve the speed and accuracy of object detection. Kumar and Singh [6] address the challenge of detecting defects by introducing a two-stage object detection network based on region proposal network (RPN) for PCB defect detection. This type of two-stage detection algorithm presents several challenges, including limited inference speed and susceptibility to interference from background noise. In contrast, single-stage object detection algorithms combine feature extraction with prediction frame positioning, enabling the direct determination of target categories and detection frame positions. Examples include SSD [7] and YOLO [8]. Kumar and Singh [9] leveraged the SSD architecture for PCB defect detection, providing a detailed account of the training process, evaluation metrics, and the effectiveness of SSD in identifying various defects, including soldering issues and component misplacements. Liao et al. [10] introduced a cost-effective detection model based on YOLOv4, which improves the activation function in the backbone and neck prediction networks, reducing both the model size and the number of parameters. Qing et al. [11] proposed an improved PCB component detection model based on YOLOv8, which utilizes the C2Focal module as the core of the backbone network, combining fine-grained local features with coarse-grained global features. Chen et al. [12] introduced the W-YOLOv8 model, which uses a gradient gain allocation strategy with a dynamic non-monotonic focusing mechanism to direct the model’s attention to anchor frames of ordinary mass, improving overall performance. While these studies have enhanced detection accuracy, a balance between detection performance and model size remains elusive.

In response to the challenges posed by small target sizes, missed detections, and excessive model parameters in PCB surface detection, this study introduces an enhanced detection model based on YOLOv8, called DVCW-YOLO. This model significantly mitigates parameter redundancy and model complexity through the integration of depthwise convolution (DWConv), VOVGSCSP modules, a self-constructed C2fCBAM module, and a weighted intersection over union (WIoU) loss function. The specific contributions of this work are outlined as follows:

1.

The C2fCBAM module is developed by synergistically combining the C2f architecture with the convolutional block attention module (CBAM) mechanism. This integration enables the detection model to effectively prioritize critical regions by enhancing salient features while attenuating less relevant ones. Consequently, a novel backbone network is established to facilitate the extraction of key features.

2.

A neck network is designed by incorporating the VOVGSCSP module and DWConv. The implementation of the WIoU loss function during training serves to alleviate the adverse effects of low-quality and extreme samples, thereby improving model accuracy, accelerating convergence, and reducing the high rate of missed detections commonly associated with PCB defect identification.

3.

We augmented the dataset and executed comparative experiments involving nine detection algorithms, which substantiate that the proposed DVCW-YOLO consistently outperforms its counterparts in most cases.

The main contributions of this paper are outlined as follows. Section 2 introduces the architecture of the proposed DVCW-YOLO defect detection model. Section 3 presents the dataset construction, including data presentation, augmentation, and defect distribution characteristics. Section 4 details the experimental setup, parameter settings, detection results, and analysis. Finally, Section 5 concludes the paper with a summary of key findings.

2. Methodology

To achieve accurate PCB defect detection, we integrate the YOLOv8n framework with a self-designed C2fCBAM module, the VOVGSCSP module and DWConv to enhance detection performance. The following section details these approaches and their construction to address the challenges of PCB defect detection.

2.1. Proposed DVCW-YOLO Network

YOLOv8 [13] is a widely utilized detector in the field of target detection. It incorporates the cross stage partial (CSP) structure within backbone network, along with the spatial pyramid pooling with factorized convolutions (SPPF) module at its output stage. The neck network features a bidirectional path aggregation network (PAN-FPN) for efficient feature integration [14]. However, the backbone and neck networks of YOLOv8 exhibit several limitations, including the restricted feature extraction capability of C2f module, high computational complexity, and a substantial number of standard convolution parameters [15]. Additionally, the complete intersection over union (CIoU) loss function does not adequately account for the characteristics of the background region, which may lead to the neglect of surrounding contextual information, consequently diminishing detection performance [16].

To address these challenges, including the large model size in fire detection tasks and the difficulty in balancing detection accuracy with model efficiency, this paper achieves several improvements to the backbone network, neck network, and loss function based on the YOLOv8n architecture, and develops the DVCW-YOLO detection model. The overall architecture of the DVCW-YOLO model is illustrated in Figure 1. All standard convolutions in both the backbone and neck networks are replaced with the more lightweight DWConv. Additionally, we introduce the C2fCBAM module in the backbone network, which incorporates the CBAM attention mechanism at both the input and output ends of the C2f module. This new module replaces the original C2f module, enhancing the model’s ability to extract features from key areas by emphasizing salient features and suppressing less important features. Secondly, the C2f module is replaced with global structure-aware convolution (GSConv) and the more efficient VOVGSCSP module in the neck structure, reducing both the number of model parameters and the overall model complexity. Finally, the WIoU loss function is used to replace the traditional CIoU loss function. This modification ensures that the improved model focuses more on normal and high-probability samples, thereby enhancing detection accuracy.

2.2. C2fCBAM Model

When the background on the surface of the PCB is complex, it can be easily confused with the actual defect areas, leading to false detections. To address this issue and enable the detection model to focus more on key information in the input features while minimizing attention to background noise, this paper optimizes the backbone network by designing and implementing the C2fCBAM module. The CBAM attention mechanism is added to both the input and output ends of the original C2f module [17], as shown in Figure 2. This structure enables the model to prioritize critical regions by enhancing salient features while suppressing less relevant ones, thereby facilitating the differentiation between normal and defective areas. This capability is particularly beneficial in scenarios where defects are subtle or when there is substantial background interference.

The CBAM consists of two components: the channel attention module and the spatial attention module [17]. The channel attention module assigns weights to the feature channels, adaptively adjusting the importance of each channel to improve the network’s ability to capture specific features. Conversely, the spatial attention module focuses on the spatial dimensions of the feature map, enhancing the network’s capacity to perceive spatial locations by adaptively adjusting the importance of each spatial position. By incorporating CBAM, the network can better learn the distinguishing features of the input data, thereby improving its ability to differentiate between various objects and backgrounds.

2.3. Neck Network

In the context of PCB defect detection, the traditional YOLOv8n model suffers from large model parameters and limited detection accuracy. To address these limitations, this paper reduces the model parameters by integrating DWConv convolutions and VOVGSCSP modules, thereby improving detection speed. Specifically, all standard convolutions in the neck structure are replaced with lighter DWConv convolutions. Additionally, due to the redundant processing of input features in the C2f module, a significant residual redundancy issue arises. To mitigate this, the C2f module is replaced with the more efficient VOVGSCSP module, further reducing model parameters and enhancing detection speed.

2.3.1. DWConv Convolution

DWConv [18] is a specialized convolution technique, and its structure is illustrated in Figure 3. Unlike standard convolution, where all channels share a single convolutional kernel, DWConv assigns a unique convolutional kernel to each channel. This technique groups the input feature maps and applies convolution within each group independently. Assuming the current number of groups is N, the number of parameters required for DWConv is only 1/N of the parameters required for standard convolution. Consequently, the use of DWConv significantly reduces the number of model parameters, resulting in a more lightweight model.

2.3.2. VOVGSCSP Module

The VOVGSCSP module [19], in contrast to the C2f module, primarily consists of multiple GSConv and standard Conv convolutions, which are connected through residual links. The GSConv module integrates several techniques, including standard Conv, DWConv, and shuffle operations [20]. The shuffle operation enables the combination of standard Conv with DWConv by mixing features across channels, which helps to reduce computational costs. As the number of input feature channels increases, the computational cost of GSConv becomes significantly lower than that of standard Conv, approximately halving the computational burden, while still preserving the feature extraction capabilities of the original convolutional operation. This structure of the VOVGSCSP module is depicted in Figure 4. The computational costs associated with standard Conv and GSConv are outlined as follows:

$$GFLO{P}_{s1}=W\times H\times K\times K\times 1\times C_{out}$$

(1)

$$GFLO{P}_{s2}=W\times H\times K\times K\times 1\times {\displaystyle \frac{C_{out}}{2}}(C_{in}+1)$$

(2)

For a given input feature map with size $W$ ×$H$, $K$ is the size of the convolution kernel, $C_{in}$ is number of input channels, and $C_{out}$ denotes the output channels. As the number of output channels increases, the computational advantage of DWConv over standard convolution becomes more pronounced. The VOVGSCSP structure effectively leverages these properties, achieving efficient feature extraction and processing.

2.4. WIoU Loss Function

The CIoU loss function [21] in the YOLOv8 model performs suboptimally when detecting low-quality samples, such as small defects, and exhibits limited generalization ability. To address these issues, the WIoU loss function [22,23] in the DVCW-YOLO model incorporates dynamic and non-monotonic focusing mechanisms. These mechanisms enhance the model’s ability to handle small objects or complex backgrounds by improving generalization and detection quality. The WIoU function assesses anchor frame quality by considering outliers in the dataset and adjusting sample weights to reduce the impact of low-quality samples.

$$L_{WIoU}=R_{WIoU}{L}_{IoU}$$

(3)

$$R_{WIoU}=\exp ({\displaystyle \frac{{(x-x_{gt})}^2+{(y-y_{gt})}^2}{(W_g^2+H_g^2)}})$$

(4)

$$L_{IoU}=1-IoU$$

(5)

Specifically, $LWIoU$ denotes the modified loss function, while ($X$, $Y$) are the horizontal and vertical coordinates of the predicted bounding box center, respectively. ($Xgt$, $Ygt$) represent the center coordinates of the true bounding box, and ($Hgt$, $Wgt$) denote the height and width of the true bounding box, respectively. The loss $RWIoU$ is computed using an exponential function, which increases the weight of anchor boxes that are closer to the true bounding box while reducing the impact of those farther away. This loss $IIoU$ complements the $IoU$ between the predicted bounding box and the true bounding box, focusing on improving the overlap quality.

3. Dataset Building

In this study, we utilize the open-source dataset from the Open Laboratory of Intelligent Robotics at Peking University, which contains six types of PCB board defects: missing hole, mouse bite, open circuit, short circuit, spurious copper, and spur. A subset of the dataset is shown in Figure 5. The dataset consists of 693 images with pixel dimensions of 2777 × 2138, with each image containing 3–6 defects. The defects occupy approximately 0.24% of the total pixel area in each image. Due to the limited size of this publicly available dataset, the small number of samples poses a challenge for effective PCB defect detection. To address this issue, we apply data augmentation techniques, including mosaic enhancement, flipping, random noise addition, and brightness adjustment, to randomly expand the dataset and improve the generalization ability of the detection model. After augmentation, the dataset contains a total of 10,688 images, and the ratio of the training set, validation set, and test set is 7:2:1. Some instances of these augmentations are shown in Figure 6.

To gain an intuitive understanding of the defect information after data augmentation, we analyzed the overall distribution, size, and quantity of defects, as depicted in Figure 7. As depicted in Figure 7a, the augmented dataset generated through random data augmentation, the number of images across the six defect categories varies, although the differences are relatively small. The number of missing holes, mouse bites, open circuits, short circuits, spurs, and spurious copper are 2883, 2935, 2798, 2805, 2915, 3000, respectively. Figure 7b illustrates the positions of the center points of all defect detection bounding boxes within the augmented images. From Figure 7b, it is evident that the defect center points of all augmentations are widely dispersed, suggesting that the defects do not follow a clear spatial pattern. Figure 7c presents the width and height of all bounding boxes in the augmented dataset. As seen in Figure 7c, the majority of the bounding boxes are concentrated near the origin, with their dimensions primarily in the range of [0.025, 0.05]. This indicates that most of the detection bounding boxes are small, with the majority of the defects being small-sized targets.

4. Experiments and Analysis of Results

In the following section, we discuss the experiment and evaluate the proposed DVCW-YOLO detection model using the augmented dataset. We performed three validation experiments to assess the performance of the model and determine key evaluation metrics and model parameters. The first experiment tested the effectiveness of the C2fCBAM module in the optimized backbone network, comparing its performance with different attention mechanisms. The second experiment conducted ablation studies on the DWConv, C2fCBAM, VOVGSCSP, and WIoU components designed for the backbone network and neck network, respectively, to evaluate the impact of each optimization on model performance. The third experiment compared the proposed DVCW-YOLO detection model with several state-of-the-art models, including Faster R-CNN, SSD, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLO-G, and YOLOv8-PCB, to demonstrate the performance advantages of the DVCW-YOLO model.

4.1. Evaluation Indicators

To comprehensively evaluate the performance of the model, six evaluation metrics are employed: precision (P) [24], recall (R) [25], mAP [26], floating point operations (FLOPs) [27], parameters [28], and FPS [29]. Precision is defined as the ratio of correctly detected positive samples to the total number of detected positives, including false positives. Recall, on the other hand, measures the ratio of correctly detected positive samples to the total number of actual positive samples, including undetected positives. The mAP represents the average precision across multiple categories and recall levels. FLOPs and the number of parameters are indicators of model complexity. FLOPs measure the computational complexity, with a higher number of FLOPs indicating a greater computational cost and increased demand for computing resources. The parameters metric refers to the number of model parameters, which provides insight into the model’s memory usage and overall size. Finally, FPS is used to assess the model’s detection speed, indicating the timeliness of the model’s performance. The formula for calculating these metrics is as follows:

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(6)

$$R={\displaystyle \frac{TP}{TP+FN}}$$

(7)

$$AP={\displaystyle \underset{0}{\overset{1}{\int }}P(R)dR}$$

(8)

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

(9)

where $TP$ (true positive) refers to the number of correctly identified positive targets. $FN$ (false negative) represents the number of positive targets that were missed or incorrectly classified. $FP$ (false positive) denotes the number of non-positive targets incorrectly identified as positive. $N$ is the total number of categories. t indicates the average detection time for each image, which is used to calculate FPS.

4.2. Experimental Environment and Parameter Settings

The experiments were conducted on an Ubuntu 18.04 platform with the following specifications: Intel (R) Xeon (R) Platinum 8255C CPU @ 2.50GHz, NVIDIA RTX 2080 Ti (11 GB GPU), PyTorch version 1.8.1, Python 3.8.0, and the VSCode-integrated development environment. Anaconda 4.13.0 was used to create a virtual environment.

Model training parameters directly influence the accuracy and performance of the model. To ensure consistency and reliability, all models were tested under the same training conditions. The specific parameter settings are as follows: learning rate (LR) = 0.001; batch size = 32; momentum factor = 0.937; and weight decay = 0.0005. The dataset was split into training, validation, and test sets at a 7:2:1 ratio. All models were initialized using transfer learning with the pre-trained weights of YOLOv8n. The training process variation curves for the bounding box loss (Box_Loss) and the categorical loss (Cls_Loss) are shown in Figure 8.

As shown in Figure 8, both the bounding box loss and classification loss of the model gradually decrease as the number of training iterations increases. In the early stages of training, the learning rate is higher, leading to a faster reduction in loss. In Figure 8a, it illustrates that the bounding box loss curve begins to flatten after 180 epochs. Similarly, Figure 8b shows that the classification loss curve decreases more slowly after 100 epochs and becomes stable around 180 epochs. Based on these observations, this study sets the total number of training epochs to 200.

4.3. Experimental Comparative Analysis

4.3.1. Comparative Experiments on Attention Mechanisms

To evaluate the detection performance of the C2fCBAM module in the YOLOv8n backbone network, comparative experiments are conducted with the gated attention module (GAM) [30] and learned spatial kernel (LSK) attention modules [31]. The detection performance of each module is analyzed, and the experimental results are presented in

Table 1. Performance comparison of three attention mechanisms.

Attention Mechanism of YOLOV8n Model Backbone Network	P/%	R/%	mAP0.5/%	FPS (f/s)	FLOPs/G	Parameters/M
NONE	96.1	95.0	95.3	41.6	8.9	3.15
+GAM	98.2	99.0	99.0	18.8	18.0	7.87
+LSK	98.2	99.0	99.1	35.7	9.3	3.35
+CBAM	98.4	99.4	99.3	41.2	8.9	3.17

.

From Table 1, it is evident that the C2fCBAM module, when utilizing the CBAM attention mechanism, demonstrates significant advantages in detection accuracy and model complexity. The mAP50 with the CBAM attention mechanism was 99.3%, which is 4.1%, 0.3%, and 0.2% higher than the mAP50 values for YOLOv8n, GAM, and LSK without an attention mechanism, respectively. Additionally, the P and R values for the CBAM attention mechanism were 98.4% and 99.4%, representing increases of 2.3%, 0.2%, and 0.2%, and 4.6%, 0.4%, and 0.4%, respectively, compared to YOLOv8n, GAM, and LSK without attention mechanisms. Moreover, adopting the CBAM attention mechanism resulted in a minimal increase of 0.02 M in model parameters, which is a negligible increment for the original model. In conclusion, the C2fCBAM module, based on the CBAM attention mechanism, effectively enhances the detection accuracy of the YOLOv8n backbone network without significantly increasing the model size, while also maintaining computational efficiency.

4.3.2. Optimize the Module Ablation Experiment

To verify the performance of each optimization operation, ablation experiments were conducted on DWConv, C2fCBAM, VOVGSCSP, and WIOU loss functions in the backbone and neck networks, respectively. Four sets of experiments were designed to evaluate different optimization strategies by incorporating various modules into the YOLOv8n model, with each experiment using identical training parameters. The experimental results are presented in

Table 2. Performance comparison of ablation experiment.

Model	DWConv	VOVGSCSP	WIOU	C2fCBAM	P/%	R/%	mAP0.5/%	FPS (f/s)	FLOPs /G	Params /M
YOLOV8N	×	×	×	×	96.1	95.0	95.3	41.6	8.9	3.15
YOLOV8N-D	√	×	×	×	98.2	99.0	99.1	42.1	7.7	2.58
YOLOV8N-DV	√	√	×	×	98.5	98.5	98.7	43.6	7.0	2.46
YOLOV8N-DVW	√	√	√	×	98.6	98.7	99.1	43.6	7.0	2.46
DVCW-YOLO	√	√	√	√	99.1	99.4	99.3	43.3	7.0	2.47

.

An analysis of Table 2 reveals several key performance improvements when comparing the YOLOv8n-D model to YOLOv8n. Specifically, the mAP of YOLOv8n-D increased by 3.9%, the FPS improved by 0.5, the model’s FLOPs decreased by 1.2 G, and the number of parameters was reduced by 0.57 M. These findings indicate that the incorporation of the DWConv convolution operation not only enhances detection accuracy but also reduces the model’s parameter count, contributing to a lighter model. When comparing YOLOv8n-DV to YOLOv8n-D, the mAP of YOLOv8n-DV slightly decreased by 0.4%, but FPS increased by 1.5, FLOPs reduced by 0.7 G, and the number of parameters decreased by 0.12 M. This suggests that while the addition of VOVGSCSP leads to a minor reduction in detection accuracy, it effectively boosts detection speed and reduces model complexity. The further comparison of YOLOv8n-DVM with YOLOv8n-DV shows a 0.4% improvement in mAP, with no changes in detection speed or model complexity. These results indicate that the YOLOv8n-DVM model, optimized with the WIOU loss function, enhances the quality of detection anchor frames, improves the model’s attention to small samples, and ultimately boosts detection accuracy. Finally, in comparison with YOLOv8n-DVW, the DVCW-YOLO model shows a 0.2% increase in mAP, a slight decrease of 0.3 in FPS, no change in FLOPs, and a marginal increase of 0.01 M in parameters. This suggests that the C2fCBAM structure used in DVCW-YOLO effectively improves detection accuracy, with only a minimal increase in model size.

To further analyze the impact of different modules on model performance, the trends of the detection loss value and mAP for each detection model were compared across training iterations, as shown in Figure 9. Under the same threshold, the DVCW-YOLO model outperforms the other models. As seen in Figure 9a, after 50 iterations, the loss curves for all models fluctuate within a small range, but the DVCW-YOLO model consistently exhibits the lowest loss, indicating that the model’s predictions are closer to the actual labels in the training data. Additionally, the DVCW-YOLO model and the YOLOv8n-DVW model converge significantly faster than YOLOv8n-DV, YOLOv8n-D, and YOLOv8n, suggesting that the C2fCBAM and WIOU loss functions facilitate optimization efficiency. Figure 9b illustrates that the mAP value of DVCW-YOLO model converges more quickly than the other four models and tends to stabilize after training 50 epochs. This confirms that the DVCW-YOLO model effectively performs PCB defect detection and exhibits clear overall performance advantages.

4.3.3. Performance Comparison of Different Detection Algorithms

To further assess the object detection performance of the DVCW-YOLO model proposed in this study, it was compared with several state-of-the-art models, including Faster R-CNN, SSD, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLO-G [32], and YOLOv8-PCB [33]. The detailed comparison results are presented in

Table 3. Performance comparative results of different models.

Model	mAP0.5/a%	FPS (f/s)	FLOPs/G
Faster-RCNN	89.1	1.2	129
SSD	76.8	23.3	31.5
YOLOv3-tiny	79.6	26.7	26.4
YOLOv4-tiny	85.9	30.1	21.2
YOLOv5s	93.6	32.1	16.4
YOLOv7	94.6	32.5	9.8
YOLO-G [32]	99.0	-	13.2
YOLOv8-PCB [33]	98.37	8.8	-
YOLOv8n	95.3	41.6	8.9
DVCW-YOLO	99.3	43.3	7.0

.

The DVCW-YOLO model achieved the highest mAP@0.5 of 99.3% among all the models evaluated. This represents improvements of 11.4%, 29.2%, 24.7%, 15.5%, 6%, 4.9%, 0.3%, 0.9%, and 4.1% over faster R-CNN, SSD, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLO-G, and YOLOv8n, respectively. In addition to its superior accuracy, the DVCW-YOLO model also demonstrates a remarkable efficiency, with a FLOPs value of only 7.0 G. In comparison, it shows a significant advantage in computational efficiency, reducing FLOPs by 122 G, 24.5 G, 19.4 G, 14.2 G, 9.4 G, 2.8 G, and 6.2 G over faster R-CNN, SSD, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLO-G, and YOLOv8n, respectively. DVCW-YOLO also had significant advantage in detection speed, and its FPS value was greater by 42.1, 20, 16.6, 13.2, 11.2, 10.8, 34.5, 1.7 than those of faster R-CNN, SSD, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLOv8-PCB and YOLOv8ns. Overall, the results indicate that the DVCW-YOLO model offers clear advantages in both detection accuracy and speed compared to other models, demonstrating its superior performance in object detection tasks.

4.3.4. The Analysis of Detection Results

To further analyze the detection effect of proposed DVCW-YOLO model, we classify the detection results for various defects in the test set, including missing holes, open circuits, short circuits, and spurious copper. The detection outcomes of the DVCW-YOLO model and the original YOLOv8n model are presented in Figure 10 and Figure 11.

As illustrated in Figure 10a–c and Figure 11a–c, the DVCW-YOLO model shows higher confidence in detecting defects such as PCB missing holes, open circuits, and short circuits compared to the YOLOv8n model. The mentioned four defects are detectable using the Yolov8n with a confidence score around 0.8. The use of DVCW-YOLO can not only effectively detect these defects but can also achieve a detection confidence of about 0.9. Notably, in Figure 10d, YOLOv8n misses a defect, while in Figure 11d, the DVCW-YOLO model accurately detects all defects. These results highlight the significant improvement in detection accuracy achieved by the DVCW-YOLO model, as well as its effectiveness in reducing the missed detection rate. This makes the DVCW-YOLO model more suitable for practical PCB defect detection applications.

5. Conclusions

To tackle the challenges of balancing large model sizes and high missed detection rates that commonly afflict traditional detection methods, we propose an advanced defect detection model for PCBs based on the YOLOv8n architecture. Our approach involves replacing all conventional convolutional layers in the YOLOv8n backbone with lightweight DWConv. Additionally, we introduce the C2fCBAM module into the backbone network to enhance the extraction of salient features while reducing the influence of secondary features. This modification allows the model to focus more effectively on critical regions, thereby improving its ability to accurately characterize and extract features of the detected objects, particularly addressing the challenge of high false detection rates for small targets. In the neck network, we substitute the C2f module with a more efficient VOVGSCSP module to optimize the trade-off between model accuracy, detection speed, and overall model size. Moreover, we implement the WIoU loss function instead of the traditional CIoU loss function, which helps alleviate the impact of low-quality and extreme samples, thereby enhancing model accuracy, accelerating convergence, and reducing the high missed detection rate typically encountered in PCB defect detection. Our experimental results on self-augmented datasets indicate that the DVCW-YOLO model achieves mAP of 99.3%, along with a detection speed of 43.3 FPS.

In future work, we plan to expand the dataset to include a broader range of defect types and various inspection objects. This initiative aims to improve the generalization and applicability of the detection model, thereby providing a solid theoretical framework for real-time industrial defect detection.