Abstract: The traditional metal surface defect detection is completed by manual visual inspection. Because of the shortcomings of manual visual inspection, such as low efficiency, high missed detection rate and high labor intensity, it is difficult to meet the efficiency and accuracy requirements of metal surface defect detection. Aiming at the problem of low efficiency for manual detection of small defects on metal surfaces in industrial production process, a method for detecting small defects on metal surfaces was proposed by enhancing YOLOv7 algorithm. Firstly, a dataset containing five types of small defects on metal surfaces was established. Then, a diffusion convolution was designed. The spacing of the feature points in the convolution kernel was enlarged using the step size to expand the receptive field of the convolution layer. Then, a directional attention module was designed. After the input feature map was segmented in horizontal and vertical directions, feature extraction was performed on the feature map. The attention mechanism was introduced in the channel dimension. According to the weights of the channels, the readjustment of the number of output channels was completed to augment the positional awareness for target of small defects in YOLOv7. Finally, the results of target detection from different algorithms on the dataset of small defects on metal surfaces were investigated, and ablation experiments were designed to carry out the performance analysis of the different improved strategy. The experimental results show that, under the same training strategy, comparing with the traditional YOLOv7 algorithm model, the detection efficiency of the improved YOLOv7 algorithm is 91 fps for small defects, and the average detection accuracy is 88.0%, which is 3.6% higher than that of the original model. It can be seen that this method can accurately detect small defects on metal surfaces under complex backgrounds in actual production.
Key words: defect detection efficiency and accuracy; improved YOLOv7 algorithm; deep learning; diffusion convolution; attention mechanism; convolutional neural network