Fabric defect detection based on separate convolutional UNet

Defect detection in the textile industry is an important and demanding task. Traditional methods rely on manual inspection, which is costly and damaging to the fabric. The deep learning methods based on semantic segmentation network simply and efficiently implement the fabric defect detection with high accuracy. In this paper, we proposed a Separation Convolution UNet (SCUNet) combined with convolutional down sampling, depth-separable convolution and cross-parallel ratio loss function(IoU Loss), and the number of parameters is only 4.27 M (Million). The location detection of fabric defects is performed by extracting surface features in fabric pictures. We selected a dataset containing 106 fabric grayscale images and performed preprocessing including image cutting and data enhancement. We tested the SCUNet with four metrics on the AITEX dataset, and the results showed that the accuracy, recall, specificity and mIoU are 98.01%, 96.86%, 98.07%, and 34.32%, respectively.