Monochromatic Fabric Defect Detection And Recognition Based On Deep Convolutional Neural Network

Textile industry is a mainstay industry in the national economy and an irreplaceable livelihood industry of our country for a long time.Moreover,it takes a leading position internationally.At present,fabric defect detection and recognition are important factors that restrict fabric production efficiency and its quality.Traditional fabric defect detection and recognition are completed manually,which results in low efficiency(the detection speed is only 10-20m/min),high false detection rate and high loss.Therefore,how to realize automation of fabric defect detection and recognition is a problem that needs to be solved urgently for cloth enterprises.The cloth defect detection and recognition system in China still remains in algorithm research based on traditional digital image processing,adopting a method of manual design features.Yet,it is hard to design and optimize its features.And at present based on Convolutional Neural Network(Convolutional Neural Network,CNN)image recognition algorithm,automatic learning is based on the data characteristics,the recognition effect is remarkable,significantly beyond the traditional recognition effect of digital image processing algorithm.Few studies have yet investigated the application of CNN to the field of cloth defect recognition both abroad and at home.Based on demands of a local cloth manufacturing enterprise,attempts an exploration of fabric defect detection and recognition system based on deep CNN.The major research contents are as follows:(1)This research investigates in a cloth manufacturing enterprise,gets familiar with its current manufacturing situation and the detection demand of cloth defects.Aiming at a large data set needed for the training of deep learning model,this research equips hardware devices including cameras and LED with production lines in this enterprise,collects cloth defect images and establishes a cloth defect image data set.(2)This research applies CNN to classifications of cloth defects.Probing into CNN and its application in the image recognition and aiming at data scale of cloth defect images,the thesis proposes simplified shallow CNN model.It can not only reduce computation amount and raise computation speed,but also prevent over-fitting.(3)This research proposes a double-network-parallel CNN training method.The simplified shallow CNN model results in decrease in recognition accuracy.In order to train CNN better and improve its performance,the thesis uses the deep network's feature map to retrieve more advanced semantic information.Besides,it constructs a double-network-parallel CNN training method by means of cross entropy loss function to maintain consistence between the shallow CNN's feature map and the deep CNN's feature map.The experiment indicates that the double-network-parallel CNN training method is superior to that of single network training,leading to 99.48%recognition accuracy.(4)This research proposes a compression algorithm based on feature map optimizing kernel parameters.Aiming at large computing resources needed for CNN,this thesis improves compression algorithm of the current model and enhances the original compression algorithm's performance.On condition of 50%compression rate,the defect recognition accuracy reach 96.99%and the processing time of per image is 10.24ms.Guaranteeing the effectiveness of detection and recognition,our CNN model can operate fast on personal computers,mobile devices and embedded devices to reduce the model's hardware cost greatly.(5)In combination of software development techniques including multi-thread technique and Qt application development framework,this research deploys our findings in personal computers.The image segmentation method is used to identify each small block,and then the depth first search algorithm is used to integrate the recognition results and thus accomplishes the whole system.This thesis tests the system empirically and compare it to manual testing results.It is found that maximum absolute deviation of cloth quality grading is merely 5.2%and its real-time processing speed is 39.5m/minutes,which is twice or three times as fast as the manual testing speed,which can replace the manual method to cloth detection.