Research On Recognition Methods For Weft-knitted Fabric Structure

A crucial stage in the design and manufacture of knitted products is analyzing the organizational structure of knitted fabrics to identify the kind of sample organization.The production of knitted goods can be made more intelligent by automating the process of determining the organizational structure of knit materials using computer vision techniques.Traditional fabric image recognition methods often have poor generalization,low recognition efficiency,and limited accuracy.This paper studies the classification and segmentation tasks of weft-knitted fabric structures using deep learning methods to handle this challenge,and it builds an efficient and accurate weft-knitted fabric recognition system.First,to address different weft-knitted fabric structure recognition tasks,two weftknitted fabric structure image datasets were constructed.Based on the different number of structure types contained in the images of the to-be-recognized knitted fabric samples,the datasets can be divided into a weft-knitted fabric classification dataset with images of single structure types,and a weft-knitted fabric segmentation dataset with images containing multiple structure types.Both datasets cover 9 different structure types of weft-knitted fabrics.Second,due to the similarity of the single-sided appearance of some weft-knitted fabric structures,it is difficult to determine their structure types based on the structural features of their single-sided images.To address this issue,a lightweight dual-branch classification network was proposed.The network first extracts feature from the doublesided images of the knitted fabric input in parallel through two weight-shared subnetworks,and then integrates the structural features of the double-sided images in a serial manner,thus accurately determining the structure type of the weft-knitted fabric.The network adopts a lightweight design and enhances the ability to focus on important feature information by adding attention mechanisms,achieving efficient end-to-end weft-knitted fabric structure classification.Experiments show that the classification accuracy of this method on the test set is as high as 99.51%.Third,for the recognition task where a single weft-knitted fabric sample image contains multiple structures,an improved encoder-decoder semantic segmentation network was proposed.By using Mobie Net V2 with dilated convolutions as the backbone network and incorporating the proposed A-DASPP module into the encoding stage,the network’s multi-scale feature extraction capability is enhanced.The DUpsampling upsampling method is used to optimize the network’s ability to restore high-resolution features in the decoding stage.Experiments show that the improved network achieves better segmentation results for weft-knitted fabric structures,with an m Io U of 84.33% on the test set,and the model’s parameter count is only 4.41 M.Finally,based on the above two recognition algorithms,a weft-knitted fabric structure recognition system was developed.The software’s various modules were designed with a focus on practicality,intuitively presenting the image acquisition process and the classification and segmentation results of weft-knitted fabric structure images.The system can achieve multi-task recognition of weft-knitted fabric structures,providing a reference example for the implementation of recognition algorithms in the textile industry.