The Research Of Fabric Surface Defects Detection And Recognition Algorithms Based On Texture Characteristics

Fabric defect detection is one of the most significant procedures for quality control in textile manufacturing industry. The manual detection is time-consuming, labor-intensive and devoid of consistency and reliability due to many subjective factors. With the development of computer vision, it has been applied widely in industrial surface detection. And developing detection algorithm with superior performance has been key problem in the computer vision system. In this thesis, we research on the current detection algorithm, and propose a novel fast and efficient fabric defect detection and recognition algorithm. Our work can be classified into the following parts:Firstly, we adopt histogram equalization to enhance image contrast. During the process of somoothing and denoising, an improved Top-Hat operator is proposed, which solves that the conventional Top-Hat operator is only effective at the structure and objectives which is smaller than itself. It firstly selects a structure whose size is slightly larger than the defect, and performs opening operation, which can generate an image similar to background; then we select a structure element whose size is slightly smaller than the defect, and perform opening operator, which can also generate an image which preserve the information of defect. In the end, we get the difference between the two results.Taking into account that a few number of fabric images which contain defects, we should develop a method which can determine whether the image contains defects. Firstly, we propose a rapid detection method based on the comparison between the energy value and histogram. And we propose a novel detection algorithm based on improved LBP and SVM. Firstly, the image is partitioned into image blocks, and then improved LBP features are extracted. We can determine whether an image block contains defect based on SVM classifier. For the images which contain defect, we use auto-correlation and FFT to find the structure of the repeating unit of fabric, Based on the fact, the defect is detected by the method of expansion and corrosion in mathematical morphology, and perform opening operator to further reduce noise affection. Then, we extract the shape features, texture features from the segmented image, and apply BP neural network to classify the fabric defect.The simulation results demonstrate that our proposed method can determine whether an image contain fabric defect with rapid speed, and can segment the fabric defect. The extracted feature can represent defect, which get the better classification results using BP neural network.