An Automatic Fabric Defect Detector Using an Efficient Multi-scale Network

Efficient and accurate fabric defect detection can be beneficial to enhance the competitiveness of enterprises. Aiming at fabric defects with multi-scale characteristics, an Efficient Multi-scale Detector (EMSD) is proposed in this paper. Specifically, by combining Self-calibrated Convolution (SCConv) and Ghost Convolution (GhostConv), a novel feature extraction network is proposed to extract low-level spatial feature maps more accurately and efficiently. Then, a Dense-connected Spatial Pyramid Pooling - Fast (DCSPPF) module is designed to integrate local and global information of low-level spatial feature maps in a way that reduces the loss of defect information. Further, a feature fusion network is constructed to extract high-level semantic feature maps and integrate them with low-level spatial feature maps by skip connections to guide defects localization. Finally, three defect feature maps of different scales are sent into detection heads for large, medium and small defects detection respectively. Experiments are conducted on public Tianchi dataset and TILDA dataset to evaluate the effectiveness of EMSD. The results show that EMSD significantly outperforms all its variants and previous works with a more lightweight network architecture, and has better fabric defect detection capability.