A Deep Self-Attention Residual Convolutional Neural Network for Automatic Recognition of Aluminum Profile Surface Defects

Due to the good properties and irreplaceable effects, aluminum profiles are widely used in various industrial products. However, complicated manufacturing environments usually lead to surface defects of aluminum profiles, which have negative impacts on product quality. Therefore, it is significant to develop an accurate and efficient aluminum profile surface defect recognition method. In this paper, a deep self-attention residual convolutional neural network (DSARNet) is proposed for automatic surface defect recognition of aluminum profiles. To ensure the information flow, the DSARNet adopts a ResNet as the backbone. Moreover, a kind of self-attention (SA) mechanism is introduced to force the DSARNet to aggregate global depth features. The experiment results show that the accuracy, Fl score, precision, recall and area under curve (AUC) of DSARNet reach 91.99%, 91.87%, 93.47%, 90.89% and 99.88% respectively. In addition, the DSARNet can process 462 images per second, which indicates that deep neural networks have broad application prospects in defect detection for metal materials.