Research On Fabric Defect Detection Algorithm And System Using Deep Learning

Fabric defects are the key factors affecting the quality of fabric.The traditional fabric defect detection methods rely on people,which is more and more difficult to adapt to the actual production efficiency.At present,there are some traditional methods which can not meet the requirements of high accuracy and real-time performance because of the variety and complexity of texture.Nowadays the deep learning technology has a strong advantage in image processing.In this thesis,the deep learning and support vector machine are applied to the problem of fabric(non-woven and woven)defect detection based on the practical application.To the non-woven fabric,the problem of defect detection is transformed into the problem of image classification after segmentation because of its relatively simple texture.The problem is solved by using the convolution neural network for feature extraction with the support vector machine for classifing,which ensure the high accuracy.But it has the problem of time-consuming.To solve it,a pre-classification model which is based on the texture features values of gray-level co-occurrence matrix and support vector machine is designed before the convolution neural network.The pre-classification model and the back final classification model make up a multi-level detection algorithm.Furthermore,the computations of the gray level co-occurrence matrix and its texture feature values are put on the GPU based on CUDA architecture in order to save computing time to the limit.To the woven fabric,applying directly the multi-level detection algorithm for the non-woven fabric to it results out not ideal,the main reason of which is that the complexity of textile fabric textures determine the accuracy of the pre-classification model is not ideal.The idea of Faster R-CNN is used in this thesis,in which the detection problem is transformed into the determination of candidate regions that may contain defects and the classification and regression of candidate regions.It uses two convolution neural networks that are used to determine the candidate regions and classify and regress the candidate regions respectively.Experiments show that the two algorithms for the corresponding type of fabric do well in both the accuracy and the real-time performance.Lastly,the thesis briefly introduces the structure and implementation of the fabric defect detection system from two aspects of hardware and software.Besides,all works of this thesis are summarized,and some efficient improvement solution has put forward at the end of the thesis.