Research on defect detection combining deformable convolution and self-attention

Abstract:
The surface defect detection of ceramic tiles has always been an important application direction of deep learning. Since the surface of ceramic tiles are mostly small defects, using methods based on traditional methods or ordinary deep neural networks to detect such small defects cannot be well detected. This paper proposes a defect detection network that combines deformable convolution and self-attention: FDTR (flaws detection network with DCN and Transformer). The network can learn the connections between pixels, which improves the ability to detect tiny types of defects. At the same time, because the direction of the flaws is relatively random, a deformable convolution module is added at the end of the feature extraction network, so that the network can not only extract those regular-shaped flaws, but also well extract flaw features with various rotation angles. In addition, because there are some high-resolution images with a size of 8000*6000 in this data set, slicing and splicing of pictures and sample balancing techniques are added to the network to improve the robustness of the network. Finally, the SOTA effect is achieved on the tile defect detection data set.