Deep Learning Based Defect Detection Research on Printed Circuit Boards

The attribute extraction network of YOLOv8 represents an enhancement upon YOLOv5. Rather than employing the Focus component for parameter reduction, YOLOv8 utilizes a 2D convolution with a stride of 2 and a kernel size of 3 for expanding channels from the initial input. It also introduces a novel convolutional unit, C2f, which takes the place of the C3 unit in YOLOv5, while maintaining the SPPF pooling element.

The C2f module is still constructed based on the CSPNet (Cross Stage Partial Network) idea, and the architecture is depicted in Figure 2. Upon inputing the attributes into the C2f unit, the 1x1 convolution will be used for channel integration, and then the feature tensor will be sliced into two parts according to the channel Split, one of which will enter the Bottleneck block to further extract the features, and the other part will be spliced with the features processed by the Bottleneck block according to the channel. This configuration augments the ability of the convolutional neural network to extract features and minimizes the time spent on memory access. The Bottleneck is illustrated in Figure 3, which utilizes the residual idea, where the original input is convolved twice to extract features and then pointwise added to the original input. One part of the output of the Bottleneck continues to be used for the Bottleneck operation, and the other part is spliced with the half after Split for per-channel splicing. The advantage of the residual idea is that it not only preserves the basic features of the original input, but also avoids the problem of vanishing gradients.

Figure 2.

Figure 3.

SPPF (Spatial Pyramid Pooling Fast) is an improved version of SPP, the structure is able to extract features at different scales of the object as SPP, enrich the feature information of the output layer of this feature, and present the same effect, while the time consumed is half of SPP, the structure is shown in Figure 4. The main process is: after the input features pass through the 1x1 convolutional integration channel, the output is copied in two copies, one for pooling kernel of 5 for maximum pooling, and the other is involved in the splicing with the output of the pooling layer. The pooled feature output is also copied into two copies, one to participate in the splicing of other pooled outputs, and the other to continue the maximum pooling, repeat this step three times, a total of four copies of the output features, four copies of the output features are spliced according to the channel, and the channel information is integrated using the 1x1 convolution.

Figure 4.

The Neck segment of YOLOv8 predominantly utilizes an enhanced version of the PANet (Path Aggregation Network) design, which builds upon the FPN (Feature Pyramid Networks). This structure involves upsampling the feature layers of smaller dimensions and fusing them with feature layers of larger dimensions, directly outputting the FPN to enrich the feature information contained in the large-dimensional feature maps. On the other hand, PANet further downsamples the fused large-dimensional features, merges them again with the feature maps of smaller dimensions, enriching contextual information, and enhancing the expression capability of the smaller-dimensional features. In contrast to YOLOv5, YOLOv8 employs the C2f unit in place of the C3 unit, and integrates channel-wise before upsampling the feature maps of smaller dimensions by removing the 1x1 convolution.

The Head component of YOLOv8 has transitioned from the initial coupled head design to the present prevalent decoupled head architecture. This decoupled head represents a standard design in object detection, tasked with deriving target location and class information from the detection network's feature map.

Specifically, decoupling the head involves separating the main part of the neural network model from the classifier part for training. The advantage of this design is the flexibility to modify and replace the classifier without altering the backbone network. Moreover, the decoupled head efficiently diminishes the quantity of parameters and the computational burden, which enhances the model's capacity for generalization and robustness, all while maintaining the backbone network architecture.

Within the YOLOv8 decoupled head structure, the prior Obj branch has been eliminated. Now, only separate branches for classification and regression persist. The regression branch employs an integral form derived from the Distribution Focal Loss concept. It is important to highlight that the channel counts in the classification and regression branches of the decoupled head may differ.

The loss function is a critical element in the training of algorithmic models; an appropriately designed loss function can lead to quicker model convergence and enhanced robustness. The loss function utilized by YOLOv8 comprises a combination of classification loss VFL (Varifocal Loss) and regression loss CIOU + DFL (Distribution Focal Loss).

The classification loss VFL is shown in Equation 1, where q represents the overlap and concurrent alignment between the projected coordinate system and the actual coordinate system, p is the softmax output value of the category, α considers the spectrum of values [0,1], and γ considers the spectrum of values [0,5]. Since there are too few positive samples during training, reducing the Loss contribution of negative samples makes the model more inclined to the training of high quality positive samples. (1) VFL(p,q)={−q(qlog(p)+(1−q)log(1−p))q>0−αprlog(1−p)q=0 VFL(p,q) = \left\{{\matrix{{- q(qlog(p) + (1 - q)\log (1 - p))} & {q > 0} \cr {- \alpha {p^r}\log (1 - p)} & {q = 0} \cr}} \right.

The regression loss L_reg mainly calculated using the summation of the CIOU loss L_CIOU and the DFL loss is depicted in Equation 2, the L_CIOU mathematical expression is illustrated in Equation 3, and the DFL mathematical expression is illustrated in Equation 6.

In Equation 3, IOU(A, B) signifies the degree of overlap and concurrent alignment between the actual and estimated coordinate systems, ρ²(b, b^gt) notes the geometric distance measured in Euclidean space between the central points of the forecasted and actual coordinate systems, c is the diagonal distance of the outer rectangle containing the real and predicted target frames, and α is the coefficient used for balancing the ratio as shown in Equation 4. v is the distance between the height-width ratio wgthgt {{{w^{gt}}} \over {{h^{gt}}}} of the real coordinate frame and the height-width ratio wh {w \over h} of the predicted coordinate frame, used to measure the height-width scale consistency as shown in Equation 5. In Equation 6, y denotes the true label value, y_i and y_i+1 denote the two closest values of y, respectively, and S_i and S_i+1 correspond to the probabilities of the two values.

The experimental environment is shown in Table I

The dataset in question is a PCB (printed circuit board) defect collection made available by the Open Lab of Peking University. The types of defects are missing hole, mouse bite, open circuit, short circuit, spur, and spurious copper, and it contains a total of 11,361 images.

When assessing single-target detection models, it's customary to utilize metrics such as accuracy and recall to gauge the model's detection efficacy. The accuracy rate is defined as the ratio of correctly identified actual samples to the total number of samples, while the recall rate represents the ratio of correctly identified actual positive samples to the total number of actual positive samples. The precise formulations for the accuracy rate P and the recall rate R are depicted in Equations 7 and 8, respectively.

Within this evaluative measure, TP signifies the count of positive samples rightly identified, FP denotes the quantity of negative samples erroneously labeled as positive, and FN indicates the number of positive samples mistakenly classified as negative. For a detection model, it is often desirable to have higher precision and recall, but it is often the case that a rise in one metric causes a fall in the other. In order to comprehensively assess the performance metric of a detection model, researchers introduce the P-R curve. The P-R curve is a curve that describes the change in the relationship between the model's accuracy and recall, and by determining the area beneath the curve, it can intuitively reflect the model's goodness or badness.

In the assessment of multi-target detection models, the metric often employed to gauge the comprehensive efficacy is the Mean Average Precision (mAP). This metric encapsulates the model's detection acuity across all categories. The mAP score is derived from the average of the Average Precision (AP) for each category. It mirrors the precision in detecting individual targets and is a function of the model's precision and recall rates. The underlying computational expressions are illustrated in Equations 9 and 10. (9) AP=∫01PRdR {\rm{AP}} = \int_0^1 {PRdR} (10) mAP=1cΣci∈CAPci {\rm{mAP}} = {1 \over c}{\Sigma _{{c_i} \in C}}{AP}_{{c_i}}

To validate the detection capabilities of the models presented in this study, a comparative analysis with contemporary mainstream object detection methodologies was conducted. The training of all models was carried out utilizing transfer learning techniques, with the COCO2017 dataset serving as the pre-training dataset. The outcomes of these experiments are documented in Table II.

The model introduced in this paper achieves superior overall performance regarding detection accuracy and speed, with a mAP of 92.3 and a detection rate of 157.2 frames per second (FPS), while only having 28.5 million parameters. Although the two-stage approach, Faster R-CNN, offers higher detection accuracy due to its maximum input image resolution, its large parameter count and slower computation make it less practical for real-world applications. In contrast, single-stage models like YOLOv5 and YOLOv7 still trail behind YOLOv8 in both accuracy and speed. These findings are presented visually in Figures 5 and 6.

Figure 5.

Figure 6.