Research On Textile Fabrics Defect Detection Algorithm And System Based On Lightweight Convolutional Neural Network

The defect is the main factor that affects the quality of textile fabric.The traditional defect detection based on machine vision uses the method of artificial feature extraction with not high detection accuracy.In recent years,a defect detection method based on deep neural network has been developed,and the detection accuracy has been improved.These large networks can only be emulated on high-performance servers,but cannot used on industrially slow embedded devices.In order to meet industry demands,a two-level lightweight convolutional neural network was designed on the embedded TX2 platform for textile fabric defect detection in this paper.At the first level,the textile fabric defect pre-classification was conducted to filter the normal textile fabric.The task undertaken in this level was simple.In this paper,based on the lightweight network Mobile Net V2 in the object detection field,the improved design was conducted,and the precision was improved by introducing Mish activation function and Cutmix data enhancement method;the speed was increased by combining the convolution layer with the batch normalization layer,and adjusting the network tail structure.The experiment showed that the precision and speed of the improved network were increased.In order to further improve the speed,the improved model was used as the benchmack model,and the GAN was designed,and residual block and channel pruning was conducted on it,showing that the model speed was increased after pruning.At the second level,the textile fabric defect final classification and localization was conducted.The task at this level was relatively complex.In this paper,based on YOLOV3,a large network in the field of object detection,the backbone network used cross-stage local network to transform Darknet53,making the model more lightweight;in view of the large difference in the size of textile fabric defects,a multi-branch dense connection module was designed in the detection network to enhance feature extraction;in order to strengthen the detection ability of small size defects,the features of the shallower layer were selected in the feature pyramid fusion;in order to increase the speed,the convolution layer is combined with the batch normalization layer.Experiments showed that the accuracy and speed of the improved model were increased,but the speed was still relatively slow.Finally,the improved model was used as the benchmark model to carry out the mixed pruning of channels,residual blocks and branches.The experiment showed that the speed of the lightweight model is significantly increased after pruning.The above two level were merged to realize the final detection system.The first level with a faster speed filtered most of the normal textile fabric,and then at the second level,the specific classification and positioning of the remaining small amount of defective textile fabric were conducted.The experiment showed that the final system's overall speed was higher basically without sacrificing the accuracy compared with the only use of the second level for defect detection.Finally,the design and implementation of the textile fabric defect detection system were finished.