Research on fabric micro defect detection method based on deep and shallow layer feature fusion

To solve the problem of confusion between micro defect features and background redundancy features in the complex background of circular weft woven fabrics, a micro defect detection model based on deep and shallow feature fusion is proposed in this paper. The model is based on the network framework of SSD target detection model, and a double-attention-driven skip residual network is introduced into the feature extraction part. The convolution of large receptive field and the maximum pooling layer are improved to make the model retain more features of small defects. At the same time, in order to further alleviate the confusion between background redundancy features and micro defect features, a deep and shallow layer feature fusion method is designed to endow the detail information of micro defect in the shallow layer with the deep layer, and then endow the semantic information of the deep layer with the shallow layer, so as to enhance the micro defect features and improve the detection rate and location accuracy of micro defect. The model testing experiment was carried out with the defect data set of circular weft machine. The improved skip connection network and deep and shallow layer feature fusion method proposed in this paper increase the average accuracy of the model by 19.25%, and finally reach 81.45%, which proves the effectiveness of the proposed model in the small defect detection task of circular weft machine.