YOLOv8-DEE: a high-precision model for printed circuit board defect detection

Traditional machine learning based methods for PCB defect detection

PCB defect detection algorithms based on machine learning typically integrate image processing techniques, such as template-based methods (Zhou et al., 2023). These methods employ image matching or subtraction to identify defect characteristics, which are then analyzed using machine learning algorithms (Li et al., 2020). Crispin & Rankov (2007) propose a method for automated PCB component inspection using a genetic algorithm with template matching. This approach tackles challenges like locating and identifying multiple components on a PCB image. It employs a normalized cross-correlation for template matching and a genetic algorithm to optimize the search process, reducing computational cost compared to exhaustive search. The results demonstrate the effectiveness of this method in detecting component placement errors on PCBs. Further, Wu et al. (2013) present a two-stage system for classifying solder joints in electronic devices. First, features are extracted from images, likely including color and shape properties. Then, a Bayes classifier efficiently separates good joints from defective ones. Finally, for defective joints, a support vector machine differentiates between various defect types. This method improves classification accuracy and efficiency by reducing the number of features needed for defect type identification. Another method, explored by Hao et al. (2013), combines a neural network’s ability to learn complex defect patterns with a genetic algorithm’s optimization capabilities for automated solder joint inspection on printed circuit boards. This approach aims to improve the accuracy and efficiency of solder joint defect detection compared to traditional methods.

However, a major challenge with template-based methods is obtaining perfect template images. Different PCB inspection tasks require distinct template images, resulting in significant time and cost inefficiencies. Additionally, most machine learning-based PCB inspection algorithms are limited to defect classification, identifying the type of defect but not its location. This limitation arises because these algorithms function primarily as classifiers without positional capabilities.

Another significant traditional method is similarity measure. Annaby, Fouda & Rushdi (2019) propose an improved normalized cross-correlation (NCC) method for defect detection in PCBs. The authors enhance the traditional NCC by incorporating image preprocessing techniques, such as image normalization and noise reduction, to increase accuracy in detecting defects. They apply a sliding window approach to measure the similarity between the reference and test images pixel by pixel, using the enhanced NCC to identify mismatches that indicate defects. However, this method for defect detection is computationally intensive due to its pixel-by-pixel comparison and sliding window approach, making it slower for high-resolution images. It also struggles with variations in lighting or alignment, reducing its robustness in uncontrolled environments.

To address these issue, deep learning algorithms offer a promising solution. Deep learning networks can extract high-dimensional features from images, enabling both defect classification and localization. Thus, the application of deep learning algorithms to PCB defect detection is likely to be the future trend.

Deep learning based methods for PCB defect detection

In recent years, deep learning-based methods, particularly CNNs, have been widely adopted for PCB defect detection (Yuan et al., 2024; Ding et al., 2019). Unlike traditional machine learning methods, CNN-based approaches automatically extract image features and streamline the preprocessing process, thereby enhancing detection accuracy and speed. Some studies are devoted to solving the complex background problem in defect detection tasks to achieve better detection performance. For example, Li et al. (2022) propose a method to addresses challenges in industrial defect detection, particularly for small defects with unclear foreground-background separation. It builds upon the YOLO object detection framework but incorporates an “expanded field of feeling” to capture a wider range of contextual information. This method also utilizes feature fusion to combine different levels of detail from the image, enhancing defect characterization. The study demonstrates that YOLO-RFF achieves high-speed and accurate detection of various industrial defects.

In order to improve the detection accuracy of the model, some studies have achieved this by modifying specific convolution modules or adding attention mechanisms in the baseline model. A model called YOLO-MBBi is proposed by Du et al. (2023) for detecting surface defects on PCBs using an improved YOLOv5 deep learning model. Authors design this model to addresses limitations of YOLOv5 in this task, such as lower accuracy and speed. It incorporates various enhancements like MBConv modules and CBAM attention for faster and more precise defect recognition. The results show significant improvement in both accuracy (over 3% higher $A{P}_{0.5}$) and processing speed (achieving near 49 FPS) compared to the original YOLOv5, making YOLO-MBBi a promising solution for real-time PCB inspection. In the road defect detection task, Wang et al. (2024) introduced C2f-Faster-EMA and SimSPPF to replace the original C2f and SPPF modules in YOLOv8s, thereby improving the representation ability of the model and achieving higher accuracy on related data sets.

Moreover, some studies improve the performance by solving the multi-scale problem among defects. For example, Wang et al. (2023) introduces a novel multi-scale module to enhance YOLO-v5 for steel surface defect detection. This new module captures features at different scales more effectively by incorporating a more robust feature pyramid network, allowing better detection of small and complex defects. It improves the model’s ability to handle varying defect sizes by extracting and merging multi-scale features through adaptive convolutional layers. This results in finer feature representation and greater accuracy, particularly for subtle defects, while maintaining real-time performance. The multi-scale module significantly outperforms the standard version in defect detection tasks on steel surfaces.

Overall framework of YOLOv8-DEE

Although YOLOv8 has demonstrated impressive results on the COCO dataset (Lin et al., 2014), recent studies have shown its generalization potential across various application domains. For instance, YOLOv8 has been successfully applied to insulator defect detection (Jiang, Hou & Wang, 2024), battery inspection (Tzelepakis et al., 2023), and steel surface inspection (Kong & You, 2024). These findings further support the network’s versatility and robustness in different industrial contexts. In PCB defect detection, defect shapes and textures, such as spurs and open circuits, are often less distinct compared to typical images. Therefore, developing a high-precision PCB defect detection model based on YOLOv8 is crucial. This study proposes a novel detection model for PCB defects, termed YOLOv8-DEE. Figure 1 illustrates the YOLOv8-DEE framework. Based on YOLOv8-L, this model incorporates an enhanced backbone network using the DSC module and an improved neck based on the EMA module. Additionally, an efficient intersection over union (EIoU) loss function for bounding box regression replaces the original complete intersection over union (CIoU) loss in the head.

Enhanced backbone with DSC module

In PCB defect detection, many areas are defect-free, and defects often exhibit only slight differences in texture, shape, and size compared to defect-free regions. This subtlety poses a significant challenge to the feature extraction capabilities of detection models. To address this issue, an improved backbone incorporating the DSC module (Chollet, 2017) is proposed in this article, which enhances the model’s feature extraction ability.

As shown in Fig. 2, DSC module comprises two components: depthwise convolution and pointwise convolution. Depthwise convolution applies a separate convolution kernel to each channel of the input feature map, then concatenates the outputs to form the final result. With N input channels, N separate convolution kernels produce N output channels, which are then concatenated to create the final output feature map with N channels. Pointwise convolution, a $1\times 1$ convolution, serves two purposes in DSC: it allows the number of output channels to be adjusted freely and performs channel fusion on the feature map produced by depthwise convolution.

The DSC module performs a convolution operation on one feature map at a time and then uses a $1\times 1$ convolution for output after fusion. This process significantly reduces computational complexity compared to standard convolution operations, with minimal loss of accuracy. Consequently, many studies adopt the DSC module as a replacement for traditional convolutional structures, making it an effective tool for developing lightweight networks (Hung, Zhang & Jiang, 2019; Wang et al., 2020).

Unlike most research, the proposed backbone in this article does not use DSC modules as replacements for traditional convolutions but integrates them directly into the backbone network. Specifically, DSC modules are added to layers P2 to P5. In layers P2 to P4, the DSC module follows the C2f module, while in P5, it is placed between the C2f and spatial pyramid pooling-fast (SPPF) modules. Experiments demonstrate that these modifications minimally increase the network’s computational load while significantly enhancing its nonlinear expression capability, thereby improving the network’s ability to fit complex functions and boosting detection performance.

Enhanced neck with EMA module

The original YOLOv8 model uses the Path Aggregation Network (PANet) structure based on the FPN architecture to enhance feature extraction. Unlike the traditional FPN, PANet fuses both top-to-bottom and bottom-to-top feature maps. However, PCB images often contain complex backgrounds with many elements irrelevant to defects, such as text, copper wires, mimeographs, and tin foil. The traditional neck structure and even PANet structure are insufficient to focus on defect features amid such cluttered background information. In other words, effectively integrating feature maps from different convolutional layers is crucial for designing enhanced feature extraction networks. To meet this challenge, an improved neck based on the efficient multi-scale attention (EMA) module (Ouyang et al., 2023) is proposed. Figure 3 presents the basic process of EMA module. The EMA module enhances interaction capabilities between different feature layers, improving the model’s detection performance. Specially, EMA employs a cross-space learning method to establish short and long information dependencies via a multi-scale parallel sub-network. It reshapes some channels into batch dimensions and designs two mutually learning sub-networks to learn input features across spaces, finally fusing their outputs. Compared to other attention modules, EMA offers superior performance with lower computational overhead.

In enhanced neck design, two EMA modules are incorporated into the upsampling stage of the original PANet. The first EMA module is positioned after the Concat module, which merges the upsampled P5 feature map with the P4 feature map. This EMA module further extracts high-dimensional feature information from the fused feature using an attention mechanism. The second EMA module is placed after another Concat module, which merges the P3 stage feature map with the upsampled features from the first EMA module. This module further extracts feature information from the upsampled features integrated with lower-level features.

EIoU loss

The original YOLOv8 framework utilizes CIoU loss and distribution focal loss (DFL) technology for bounding box regression and binary cross-entropy (BCE) for classification prediction. CIoU loss serves as the loss function for target positioning, effectively measuring the model’s accuracy in detecting the target’s position and size, thereby enhancing performance evaluation. The formula for CIoU loss is as follows:

(1) $$L_{CIoU}=1-IoU+\frac{{\rho }^2(b,b^{gt})}{c}+\alpha \upsilon$$where $IoU$ denotes intersection over union, $b$ and $b^{gt}$ denote the centroids of the predicted bounding box and the true bounding box, respectively; $\rho$ denotes the Euclidean distance between the two centroids; $c$ denotes the diagonal length of the smallest closed rectangle containing the predicted bounding box and the true bounding box; $\upsilon$ is used to quantify the consistency of the aspect ratio; and $\alpha$ is a weight function. The equations for $\alpha$ and $\upsilon$ are as follows:

(2) $$\alpha =\frac{\upsilon }{(1-IoU)+\upsilon }$$

(3) $$\upsilon =\frac{4}{{\pi }^2}\left(\arctan \frac{w_{gt}}{h_{gt}}-\arctan \frac{w}{h}\right)$$where $w_{gt}$ and $h_{gt}$ denote the width and height of the ground-truth bounding box, and $w$ and $h$ denote the width and height of the prediction bounding box, respectively.

Although CIoU loss considers both the width-height ratio of the regression box and the distance between the centroids of the true and predicted bounding boxes, it has a limitation: it only uses the width-height ratio as an influencing factor. If two bounding boxes have the same center points and width-height ratio but different dimensions, CIoU loss may produce consistent results that do not align with the regression target. This deficiency becomes more pronounced in PCB defect detection tasks, where a large detection scale can exacerbate missed detections. To address this, EIoU loss (Zhang et al., 2022) was introduced. EIoU loss separates the influencing factors of the aspect ratio of the predicted and true bounding boxes, calculating their height and width independently. EIoU loss consists of three components: IoU loss ( $L_{IoU}$), distance loss ( $L_{dis}$), and height-width loss ( $L_{asp}$), which account for the overlapping area, the distance between center points, and the aspect ratio. The height-width loss directly minimizes the difference in height and width between the predicted and true bounding boxes, enabling faster convergence and improved positioning accuracy. In addition to its use in PCB defect detection, the EIoU loss function has also been successfully applied in various other domains (Jiao & Xing, 2024; Trinh et al., 2024). The formula for EIoU loss is as follows:

(4) $$\begin{array}{rl}{L}_{EIoU}& =L_{IoU}+L_{dis}+L_{asp}\\ & =1-IoU+\frac{{\rho }^2(b,b^{gt})}{{(c_w)}^2+{(c_h)}^2}+\frac{{\rho }^2(w,w^{gt})}{{(c_w)}^2}+\frac{{\rho }^2(h,h^{gt})}{{(c_h)}^2}\end{array}$$where $c_w$ and $c_h$ denote the width and height of the minimum enclosing box of ground-truth bounding box and prediction bounding box, respectively. Figure 4 shows the schematic diagram of EIoU loss.

Experimental configuration

The experiments are conducted on a high-performance computing environment. The system is equipped with an AMD EPYC 7742 64-Core processor and an NVIDIA A100-SXM4-40GB GPU, running Ubuntu 20.04.2 as the operating system. The deep learning models are implemented using PyTorch version 1.8.1 and Python 3.8.10. For optimal hardware utilization, CUDA version 12.2 is employed to accelerate GPU computation. During the training process, the batch size is set to 32, and the models are trained for 300 epochs to ensure sufficient learning and convergence of the algorithms.

Dataset

HRIPCB dataset

The public PCB defects dataset HRIPCB (Huang et al., 2020) from Peking University contains 693 images of six defect types, with an average pixel size of $2\text{,}777\times 2\text{,}188$. Due to the large size of the original images compared to the input size of typical object detection models, images are cropped into sub-images with $640\times 640$ pixels size by using sliding window technique. To prevent target frames from being segmented during cropping, some sliding windows are adjusted appropriately. A total of 4,205 images are obtained after cropping, each containing 1–4 defects. Figure 5 introduces the six types of defects on the cropped HRIPCB dataset. The training set and test set are divided in a 9:1 ratio. The number of each defect category is shown in Table 1.

Table 1:

Number of each defect category on HRIPCB dataset.

Category	Numbers
Missing hole	748
Mouse bite	756
Open circuit	706
Short	782
Spur	741
Spurious copper	772

DOI: 10.7717/peerj-cs.2548/table-1

DeepPCB dataset

All images on DeepPCB defect dataset (Tang et al., 2019) were obtained from a linear scan CCD. The original size of the test images is about $16k\times 16k$ pixels, and then they are cropped into many sub-images of size $640\times 640$, a total of 1,500 pairs of template and tested images with annotations, including six common PCB defect types: open, short, mousebite, spur, pin-hole, spur. Figure 6 introduces the six types of defects on the DeepPCB dataset. The number of each defect category is shown in Table 2.

Table 2:

Number of each defect category on the DeepPCB dataset.

Category	Numbers
Pin-hole	1,320
Mousebite	1,748
Open	1,702
Short	1,317
Spur	1,445
Copper	1,321

DOI: 10.7717/peerj-cs.2548/table-2

Evaluation metrics

To evaluate the model’s performance, the common metrics in object detection tasks are used in this article, which include precision (P), recall (R), F1 score (F1), average precision (AP), and mean average precision (mAP). The formulas for these metrics are as follows:

(5) $$P=\frac{TP}{TP+FP}$$ (6) $$R=\frac{TP}{TP+FN}$$ (7) $$F1=\frac{2\cdot P\cdot R}{P+R}$$ (8) $$AP={\int }_0^{1}p(r)dr$$ (9) $$mAP=\frac{1}{N}\sum _{i=0}^{N}A{P}_i$$where TP represents the number of targets detected correctly; FP represents the number of targets detected incorrectly; FN represents the number of defect samples falsely detected; P represents precision; R represents recall; AP values are the area enclosed under the curves of P and R, $mAP$ represents the sum of AP values of all categories, and N represents the total number of categories.

To further evaluate the model’s accuracy at different IoU thresholds and target sizes, several performance metrics from the COCO dataset, including $A{P}_{0.5}$, $A{P}_{0.75}$, $A{P}_{0.5:0.95}$, $A{P}_S$, $A{P}_M$, and $A{P}_L$ are also used as evaluating indicators in the comparison experiments. $A{P}_{0.5}$ is the average precision for each category at an IoU threshold of 0.50, and $A{P}_{0.75}$ is the average precision at an IoU threshold of 0.75. $A{P}_{0.5:0.95}$ is the average precision across IoU thresholds from 0.50 to 0.95 in 0.05 increments. $A{P}_S$ represents the average precision for small objects (area $<32\times 32$), $A{P}_M$ for medium objects (area between $32\times 32$ and $96\times 96$), and $A{P}_L$ for large objects (area $>96\times 96$).

Experiment result

Comparison experiment on the cropped HRIPCB dataset

In addition, this study compared the performance of our model with other state-of-the-art models, including YOLOv5-L, YOLOv7-L (Wang, Bochkovskiy & Liao, 2023), YOLOv8-X, Faster-RCNN, and DETR (Carion et al., 2020) on cropped HRIPCB dataset. The comparative performance of these models, including the proposed YOLOv8-DEE, is detailed in Fig. 7. This analysis focuses on several key metrics: AP at different IoU thresholds ( $A{P}_{0.5}$, $A{P}_{0.75}$, $A{P}_{0.5:0.95}$), and average precision for small ( $A{P}_S$), medium ( $A{P}_M$), and large ( $A{P}_L$) objects. The YOLOv8-DEE model demonstrates a clear advantage across all specified metrics, underscoring its robustness and efficiency in object detection tasks across varying scales and complexities.

In AP metrics at different IoU thresholds, the proposed YOLOv8-DEE model significantly outperforms all other models, achieving an $A{P}_{0.5}$ of 97.5%. The closest competitors, YOLOv5-L and Faster-RCNN, show competitive but lower performance. Additionally, YOLOv8-DEE leads in more stringent metrics, with $A{P}_{0.75}$ at 68.6% and $A{P}_{0.5:0.95}$ at 60.4%. This indicates that YOLOv8-DEE not only identifies objects more accurately but also aligns better with the ground truth, a critical factor in applications requiring high precision. The visualization results of detection effect on these models are displayed in Fig. 8. It can be clearly seen that PCB defects only account for a small proportion of the board. YOLOv8-DEE can detect all defect types and obtain more accurate defect locations than other models.

Regarding performance across object sizes, YOLOv8-DEE demonstrates a remarkable ability to detect small objects, achieving an $A{P}_S$ of 71.4%, significantly higher than any other model. This performance is crucial for scenarios where small object detection is essential, such as PCB defect detection, medical imaging, and satellite imagery. For medium-sized objects ( $A{P}_M$), while the improvements are less dramatic, YOLOv8-DEE still matches the best of the other models, showing balanced performance across object sizes. Furthermore, for large objects, YOLOv8-DEE excels with an $A{P}_L$ of 70.2%, indicating robust performance in detecting larger objects, which is often challenging due to higher variability in appearance.

Comparison experiment on DeepPCB dataset

To further validated the performance of YOLOv8-DEE, a comparion experiment is conducted on the DeepPCB dataset. Figure 9 presents the comparison of the experimental results for various models, including YOLOv5-L, YOLOv7-L, YOLOv8-L, YOLOv8-X, Faster-RCNN, DETR, and the proposed YOLOv8-DEE. The evaluation metrics used are $A{P}_{0.5}$, $A{P}_{0.75}$, $A{P}_{0.5:0.95}$, $A{P}_S$, $A{P}_M$, and $A{P}_L$.

In terms of $A{P}_{0.5}$, the YOLOv8-DEE achieved the highest score at 98.7%, outperforming all other models. Although YOLOv8-X has a more complex feature extraction network, its detection performance is still slightly lower than the proposed model. On the other hand, all anchor-based models (YOLOv5-L, YOLOv7-L,Faster-RCNN) perform significantly worse than anchor-free models (YOLOv8-L, YOLOv8-X, DETR, YOLOv8-DEE), especially Faster-RCNN, which achieves the lowest detection accuracy on all detection metrics. This indicates that anchor-free type detection heads may be more suitable for detecting multi-scale PCB defects. The visual result of YOLOv8-DEE on DeepPCB is shown in Fig. 10.

Regarding the AP metrics of different sizes, the YOLOv8 series models have achieved a greater advantage compared to other detection models. The proposed model shows stronger performance in detecting small object defects and large object defects compared to YOLOv8-L and YOLOv8-X.