Segmentation Network for Textile Sewing Stitches

In order to solve the problems that fabric stitches are difficult to be accurately segmented due to the variety of fabric stitches, the uncertainty of the types, and the similarity between stitches and fabric texture, a textile stitches segmentation network based on multi-scale features is proposed. Firstly, the residual network with attention mechanism is used to extract the location information; and the enhanced pyramid module is used, making full use of the multi-scale features to obtain the semantic information of the pre-candidate areas. After fusion, the candidate regions of sewing thread are filtered; Finally, the segmentation of sewing is realized through full convolution network, realizing the segmentation of sewing line. On real textile sewing data sets, SewTrace, the mean average precision of the proposed network is 0.96, the computational load is 1.5 G. On the open data sets CrackForest, CRKWH100 and Koletor with similar linear characteristics, comparative analysis is carried out with other similar networks, the experimental results show that the mean average precision of the proposed network is 0.85, 0.89 and 0.89, respectively, the segmentation precision and prediction speed indicators are better than other similar networks, which shows that the proposed network can effectively improve the segmentation precision of linear objects.