Research On Small Scale Defects Detection Method Of Zipper Chain Teeth

Zipper is an indispensable commodity in our production and life.At present,China is the largest zipper manufacturer and exporter in the world.In the zipper manufacturing process,due to the different quality of raw materials and the complex manufacturing process,various quality defects are inevitable.There are many types of zipper defects,few samples and different sizes of defects.Through the statistical analysis of the data distribution of chain teeth,it is found that the overall proportion of small-scale chain teeth defects is up to 46.5%,especially the small-scale target proportion of chain teeth defects like Xiaomi is up to 90.7%.Because of its small size and limited available feature information,the detection accuracy of small-scale chain teeth defect is low and difficult.Therefore,this paper focuses on the difficulties of small-scale defect detection of zipper chain teeth.The main research contents are as follows.(1)Aiming at the problem that the expression ability of middle and shallow layer features in the backbone network of target detection is insufficient,which leads to the difficulty of small-scale chain teeth defect detection,a middle and shallow layer feature enhancement network based on light-weight auxiliary network is proposed.The middle and shallow layer features generated by the network are integrated into the middle and shallow layer features of the backbone network.The light-weight auxiliary network provides the middle and shallow layer features that can reflect the real data for the backbone network of object detection.To enhance the shallow feature expression ability of backbone network and improve the detection performance of small-scale chain teeth defects.(2)In order to solve the problem that pooling operation in spatial pyramid pooling in YOLOv3-SPP will lose the feature of small-scale chain teeth defects and provide insufficient context information,a context feature extraction module based on spatial pyramid convolution is proposed,which uses the dilated convolution with different dilation rates to replace pooling operation and fuses the context feature information of different scales,So the defects detection performance of small-scale chain teeth is improved.(3)In order to further enhance the fusion and interaction between shallow fine-grained feature information and deep high-level semantic feature information,a multi-scale feature fusion module based on bi-directional feature circulation network is proposed,which includes a bottom-up feature fusion mode and a top-down feature fusion mode to enhance the interaction and fusion between shallow and deep features,so as to improve the detection accuracy of small-scale chain teeth defects.(4)In order to meet the real-time detection requirements of industrial production lines,this paper proposes a multi-scale network model based on the cross-stage split-flow network to accelerate the model and design the model light-weight.The input feature map is divided into two parts along the channel direction.One part performs normal residual convolution operation,the other part is directly connected to the output,which not only increases the flow path of gradient in network training,but also reduces the calculation of residual convolution by half,and then improves the reasoning speed of the model by 16.6% without losing the model accuracy.The mAP of the improved model in the chain teeth data set reached 0.9994,which was 3.97%higher than YOLOv3-SPP,and the reasoning time of single image on the test data set was21.36 ms.The improved model meets the requirements of the real-time and precision of the industrial production line,and overcomes the difficulty of defects detection of small-scale chain teeth.