Fabric Defect Recognition Based On Aggregate Channel Feature And Local Visual Saliency

China's textile industry has a large production scale and occupies a leading position in the global market,but the production mode is still in the labor-intensive production stage.At present,most of the quality inspections in the production process of fabric rely on manual inspections.Not only are they inefficient,but due to the influence of external objective factors,the rate of missed inspections is relatively high,which greatly reduces the quality of the fabric.In addition,with the increase in labor costs of enterprises,the improvement of market quality requirements,and the expansion of production capacity requirements,relying solely on traditional manual testing has been unable to solve the contradiction between production efficiency,product quality and production capacity,and urgently needs to be improved through enterprise transformation and industrial upgrading its market competitiveness.The construction of a high-performance intelligent fabric defect recognition system has become one of the main direction of the current textile industry upgrade.This paper is mainly aimed at the defect detection of solid color plain weave cloth produced by double needle bed warp knitting machine.Using machine vision and deep learning technology,we developed a set of airborne real-time fabric defect recognition system.The system has been successfully used in the actual production of many Changshu textile enterprises,which can effectively reduce the labor cost of the enterprise and improve the quality of the raw materials of the grey fabric.The main research work of the paper is as follows:(1)This method takes advantage of the unidirectional movement of the fabric during the weaving process,fuse the optical flow field of the fabric calculated by the continuous images of multiple frames,and uses the improved specular elimination algorithm to compensate for the uneven illumination,and finally combines the edge detection and the maximum connected domain to determine the fabric surface range.Experimental results show that this method has the advantages of low computationalcomplexity and high positioning accuracy.(2)A fabric defect detection algorithm based on aggregate channel features is proposed.This method takes advantage of the obvious difference between the texture of flawed and normal fabrics.First,it constructs training datasets of normal and flawed fabrics,then extracts a variety of manual features,construct aggregate channel features then uses a cascade classifier based on Adaboost algorithm to learn the classification model,and finally uses a sliding window method for defect detection on the fabric area.Experimental results show that the method can achieve a detection rate of 97.3% on the data set collected by this subject.(3)A fabric defect classification algorithm based on visual saliency is proposed.First,visual saliency is used to extract the saliency features,and then the MobileNetV2-FPN structure is used to extract the depth features to improve the performance of multi-scale detection on small target data.Finally,the softcutoff loss function is used to solve the problem of small difference between defect classes and uneven distribution of data.Experiments show that the method in this paper can achieve a classification accuracy of 98.5% on the data set collected by this subject.(4)Based on the above fabric positioning,defect detection and classification algorithms,this paper builds a fabric defect monitoring system for the production process and deploys it to the industrial computer for actual measurement.The test results show that the system can significantly reduce the defective rate of the product and can realize 24-hour uninterrupted real-time detection,which is an efficient solution that can be used to replace manual inspection.