HMFCA-Net: Hierarchical multi-frequency based Channel attention net for mobile phone surface defect detection

Abstract

The surface defect detection is an important process in the production of mobile phones. To detect various mobile phone surface defects and acquire detailed features of tiny defects, this paper proposes a Hierarchical Multi-Frequency based Channel Attention Net (HMFCA-Net). In particular, an attention mechanism that uses multi-frequency information and local cross-channel interaction is proposed to represent the weighted defect features. A deformable convolution based ResNeSt network is introduced to handle various defect shapes. Besides, to overcome the extreme aspect ratio problem caused by the tiny phone surface defects, a RoI Align is introduced to decrease localization error. Experiments on the public DAGM dataset and a self-collected dataset named MPSSD shows that the proposed method achieves promising performance on defect detection task.