Surface Defect Detection Algorithm Based on Feature-Enhanced YOLO

Surface defect detection is a complicated task to achieve both specific class and precise location of each defect. Specifically for industrial scenario, realizing efficient and accuracy-satisfactory surface defect automatic detection is still a big challenge. Therefore, a surface defect detection algorithm based on feature-enhanced YOLO (FE-YOLO) for practical industrial applications is proposed in this paper. For the purpose of efficient detection, we lighten YOLO model by combining deep separable convolution and dense join. And an improved feature pyramid network is proposed to enhance the spatial location correlation for multi-scale detection layer for the sake of high accuracy. Then, a new loss function of prediction box regression is established to boost the detection accuracy under the high Intersection over Union (IoU) threshold and accelerate model convergence. To select anchor boxes of different scale feature detection layers, we propose a statistical-based k-means++ algorithm, which can improve the quality of initial anchors and accelerate the convergence of the proposed model. Two industrial surface defect datasets, NEU-DET dataset and DeepPCB dataset, are used to verify the effectiveness of the proposed FE-YOLO algorithm. Experimental results demonstrate that FE-YOLO algorithm is lightened nearly 80% compared with YOLOV4. The detection speed is better than the other state-of-the-art surface defect detection algorithms. The defects detection accuracy respectively reaches 83.9% and 98.9% for the NEU-DET dataset and DeepPCB dataset, which are better than the state-of-the-art defect detection methods. The end-to-end fast and accurate detection for industrial surface defects is realized.