Fabric Defect Detection Based On Wavelet Transform And Multiscale-multidirectional Autocorrelation Approach

According to the improve of living standard, people need to acquire high-quality textile from the textile industry. One of the most important reason of impact quality is the textile defect. In the textile industry, the defect detection of woven fabric always depends on manual work, that wastes a lots of manpower and material. Meanwhile,the efficiency of inspection are limited. Therefore, the development of system which is online, real-time and automatic is of great significance. The main subject of this study is researching defect detection algorithm and designing on-line automatic defect detection system based on machine vision.(1) In this paper, the common defects and stitching defects are researched. The defect detection algorithm based on wavelet transform, Gray-level Co-occurrence Matrix and BP neural network is proposed for the commom defects. Firstly, the wavelet transform is the first step of sample image processing, because different textiles have different structure.The appropriate decomposition scale is selected in processing of wavelet decomposition.Secondly, after wavelet decomposition, the obtain horizontal high-frequency sub-image and vertical high-frequency sub-image of four sub-image are used to extract the feature value, which is based on Gray-level Co-occurrence Matrix. Finally, the BP neural network is used to defect detection. The experiments show that the algorithm can accurately detect common defect and the detection effect is better. In the second part of the part, a method based on improvement threshold algorithm with wavelet transform and BP neural network is presented to detect five kinds of stitching defects. The wavelet transform module constitutes a step of image segmentation with the objectives of attenuating the background texture and accentuating the anomalies. Low-frequency sub-images are used to enhance defect information in function reconstruction. Therefore, low-frequency sub-image is chose to complete segmentation and obtain binary image. Feature values of binary image are extracted based on the spectrum of Fourier transform method. Classification is completed by BP neural network. The experiments prove that the proposed method has good accuracy for stitching defect detection and recognition.(2) This thesis presents a new multiscale-multidirectional autocorrelation approach fordefect detection. The normal fabric texture has three visual characteristics, such as periodicity, directionality and randomness. Defect breaks the visual characteristics and make distortion of the texture, it mainly highlight on the directional characteristics.Therefor, a three scales and three directions autocorrelation approach are proposed. Firstly,select a suitable scan window size for scanning the image. Then, achieve the multiscale-mu ltidirectional by mean downsampling and image blocks rotation. Finally, calculate the variance of the autocorrelation function of each image block, the variance are seen as the feature vector put into LVQ neural network classifier to finish test. In this paper, the proposed methods are compared with the method based on Log-Gabor, the experiments showed that the proposed algorithm for defect detection had higher accuracy and pace than the method based on Log-Gabor.A set of automatic defect detection system are designed and set up in a laboratory environment. The system combines with hardware and software to achieve on-line and real-time defect detection. Experimental results show that the proposed system is reliable,accurate, real-time and well-used in the industrial field.