Research Of Fabric Pattern Detection Based On Convolutional Neural Network

With the development of artificial intelligence technology,computer-aided detection technology plays an increasingly important role in production and life.Computer-aided detection technology can not only help textile practitioners improve the accuracy of fabric pattern classification,but also improve the efficiency of fabric pattern classification.At the same time,it can ensure the employees to obtain the supply channel of fabric fabric efficiently,and expand the sales network of fabric fabric.Therefore,how to ensure that the algorithm can still accurately classify fabric patterns in complex environment,and achieve real-time performance on mobile devices and embedded devices,and at the same time meet certain classification performance has important research significance.Thesis mainly studied in the actual production environment,how to use the real-time of fabric image fast for fabric pattern classification problems,to meet the needs of training model,with the aid of many enterprises in the industry support,complete unified private fabric pattern data set,at the same time with the depth of the convolution neural network algorithm to test the fabric pattern classification,In order to solve the problem of accuracy and efficiency of convolutional neural network algorithm in fabric pattern classification,corresponding improvement is made.The main contents are as follows.(1)The training effect of deep convolutional neural network model often benefits from the amount of data set.This paper combines the fabric pattern data set scanned and collected by enterprises in the industry and the open source fabric pattern data set to make a private fabric pattern data set using the unified image entry standard and having the corresponding relationship between the pattern and the enterprise.Through preprocessing and data enhancement,the private fabric pattern data set in this paper reaches tens of thousands of images,which initially meets the training requirements of deep convolutional neural network algorithm.Moreover,the improved Detection method of Faster RCNN in Chapter 4 reduces the input standard of image input.(2)In the fabric pattern classification algorithm based on improved residual network,the residual convolutional neural network fabric pattern under sampling method in the image classification algorithm of modified into average pooling method,at the same time under the process of samples from 1×1 convolution layer over the 3×3 convolution,experimental results show that the improved algorithm solved that 1×1 is solved by using the convolution layer down sampling may cause information loss problem,use multiple small convolution kernels instead of convolution kernels,pruning part of the structure.The problem of too many parameters and layers of residual convolutional neural network model is avoided;In the improved target detection algorithm of Faster RCNN,DIo U method is used to adjust the fabric pattern classification of target detection algorithm of anchor box evaluation mechanism,minimizes the prediction box and the normalized distance between the target frame,the improved algorithm of candidate regions pooling method and the maximum inhibition method,the experimental results show that the improved algorithm can effectively solve the practical application process,the problem of the sample source range under test,Thus the classification accuracy of fabric pattern target detection is improved indirectly.(3)The deep convolutional neural network model was trained by self-made fabric pattern data set,the algorithm running environment was set up on the server,and the mobile display software was designed to display the classification results of fabric pattern in real time,and the fabric pattern detection and classification system was designed that could be applied to the actual production scene.