Study On Computer Vision-based Automatic Detection Of Fabric Defects

The automatic detection of fabric defects based on computer vision by using adaptive wavelet transform method was studied in this thesis. The main contents covered by the thesis are the construction of adaptive wavelet basis, decomposition of textural images of fabrics using adaptive wavelets, extraction of the textural features from the subimages, detection of fabric defects according to the remarkable variation of the feature index and etc. The following are the brief introduction of the content for each chapter.In foreword the background of topic selection of this thesis is described.In chapter 1 the recent progress in the world of studies on the computer vision-based automatic inspection of fabric detects is introduced. An outline is given of the methods of image process for detecting the fabric defects both in time domain and frequency domain, and especially with the emphasis on Fourier transform and wavelet analysis. It is also mentioned in chapter 1 that there are commercial systems for automatic inspection of fabric defects produced by some developed countries being already available in the world markets. However, the work carried out in our country is just on the laboratorial level now.In chapter 2 a brief introduction of wavelet analysis method is given. Except short mention of wavelet theory review, Fourier Transform, Short Time Fourier Transform, Continue Wavelet Transform and Discrete Wavelet Transform in the chapter, the fundamental concepts of Multi-resolution Analysis and Orthogonal Wavelet Transform are mainly introduced from the point of view of engineering application, which include multi-resolution analysis, two-scale equation and multi-resolution filter, fast algorithm of orthogonal wavelet transform, construction of orthogonal wavelet basis and etc. These theoretical knowledge are very helpful for thepresent work to detect the fabric defects by computer vision.In chapter 3 the orthogonal wavelet transform, i.e. orthogonal wavelet decomposition of fabric images is introduced. After discussing the traditional pyramid Mallat wavelet decomposition, the full multi-scale wavelet decomposition and the tree-structured wavelet transform of fabric images in detail, an adaptive wavelet decomposition with one resolution level is proposed. Such a decomposition of fabric images can make both the warp and weft texture be well matched with the warp and weft subimages respectively and is possessed of many advantages, such as fast computation speed, separately processing the textural information of warp and weft, more cleared texture details and quick detection of fabric defects.In chapter 4 the method of constructing the adaptive orthonormal wavelet basis is discussed. It can be constructed by optimizing the wavelet basis to be subject to orthonormal condition and constraint condition simultaneously, i.e. by minimizing the cost function using Lagrange multipliers method. However, it will be involved in a huge computation and only a relation minima be obtained. A new method is proposed in the present work, by which the wavelet basis is subjected to orthonormal condition and constrain condition sequentially, i.e. an orthonormal wavelet database for a certain length of wavelet bases is established at first according to orthonormal condition, then take the energy of high-pass decomposed subimage to be minimum as the constraint condition to search out the optimum bases from the database by genetic algorithm and iteration algorithm, these wavelet bases are certainly adaptive to the fabric texture. The experimental results prove that the fabric adaptive wavelets are better than DB wavelets with respect to the fabric image decomposition and defect detection.In chapter 5 feature parameters extraction and window section are described in detail. Five different feature parameters such as energy, standard deviation, entropy, extreme deviation and contrast are considered to be the indexes which can multiply manifest the abnormal change of the fabric texture. The window is sectioned according to the autocorrelati...