Wavelet Transform Based Texture Analysis With Application To Defect Detection Of FPC Gold Surface

With the development of computer vision, research and application of texture analysis technology become more and more intensive and extensive. However, due to the complexity of natural textures, not a general method which is well qualified for various types of texture analysis tasks has been found now. In the majority of texture analysis methods the wavelet-based texture analysis has been developed as a more promising method in recent years. A wavelet has good time-frequency localization ability and species diversity, which makes it particularly suitable for texture image processing. This dissertation mainly studies applying some features of wavelet domain to solve the part key problem of the current texture analysis techniques with expectation of implementing efficient defect detection for FPC gold surface. The main innovative achievements are as follows:A special scheme of texture segmentation is proposed based on complementary feature extraction from wavelet packet frame decomposition sub-bands to solve such problem as energy and entropy features ignored the texture orientations and the relations among local neighborhood pixels in texture segmentation, where histogram of oriented gradient descriptor is firstly introduced into the feature extraction of wavelet packet frame decomposition sub-bands. In feature extraction stages of texture segmentations, the texture image to be segmented is firstly decomposed by wavelet packet frame, and then the appropriate sub-band coefficients are selected using the designed sub-band reservation/removing algorithm, and two type of features which include the average absolute deviation and the mean and standard deviation of histogram of oriented gradient are calculated from the selected sub-band coefficients. Fisher Linear Discriminate Analysis and texture segmentation experiments both show that the combination of these two types of features has stronger ability to distinguish textures than any single one, thus which suggest that these two types of features have a certain degree of complementarities between them. In pixel clustering stages of texture segmentations, an improved spatial fuzzy c means clustering method was designed for solving the problem of higher error pixel segmentation rates near the border of textures. This method takes the standard deviation distribution of neighborhood pixel features into account. Texture segmentation results show that this clustering method can reduce segment error rate of pixels near the texture border to some degree. Pixel-based texture segmentation is similar with pixel-based texture defect detection process, therefore key technologies in texture segmentation scheme is applied to defect detection of FPC gold surfaces and more accurate defect results is obtained. A complex local binary pattern operator (CLBP) in complex wavelet domain with simplicity and efficiency is proposed for implementing the texture image retrieval and the defect detection of FPC gold surface in consideration of the reason that the existing methods using energy parameters, generalized Gaussian distribution (GGD) model parameters and generalized gamma distribution (GΓD) model parameters in wavelet domain to represent some textures lack accuracy and real discrete wavelet transform have a shift variability and weak direction selectivity. The extracted features using the simple and efficient CLBP operator can capture local texture primitive structure property to some extent. Tasks of the texture retrieval and FPC gold surface detection are both implemented firstly by decomposing texture images using dual-tree complex wavelet, and then by treating the obtained complex coefficients with CLBP whose outputting result histogram serve as a texture feature vector, and finally by calculating symmetric Kullback-Leibler divergence. Texture retrieval and FPC gold surface defect detection experiments both show that the new method is compared with GGD and GΓD model approaches of a complex wavelet decomposition to respectively improve the search accuracy rate of 8.66% and 5.48%, to respectively improve detection accuracy rate of 8.75% and 6.25%.To further reduce the feature dimension and build more robust features a feature combination method is proposed based on CLBP operator for texture retrieval and FPC gold surface defect detection, where the reduced CLBP outputting histogram is combined with GΓD model parameters of complex wavelet coefficient amplitude to form a set of features with stronger discriminatory power. Further experiments show that this method of combing features can achieve higher retrieval and detecting accuracy than the existing variety of methods. This method has practical significance and application prospects on account of the combined feature set with characteristics of rotation invariant and low dimensions.A completely unsupervised defect detection method for FPC gold surface based on Gabor filters and Mean-shift cluster is put forward. This method is characterized by the detecting process not requiring prior knowledge of the standard FPC gold surface texture type and defect texture type. The image to be detected pass Gabor filters, morphological open operation, a Gaussian smoothing filter and dimensionality reduction operation, and then Mean-shift clustering is applied to feature data and final detecting results are obtained by binarying the clustering outputting data. Mean-shift algorithm is an unsupervised clustering method, so this process need not take the total number of background and defect texture types as parameters. Experiments show that the designed method can detect various types of defects without sensitiveness to small variation in the background texture. Finally, an automatic inspection system for FPC gold surface is designed based on a variety of defect detection algorithms for FPC gold surface. Each subsystem has been specially designed for smoothly implementing the automatic defect detection of FPC Gold surfaces according to characteristics of the captured image. The detection process with three stages for the purpose of the different detection is proposed so that users can choose detection stages and detection algorithms according to different practical requirements for their detection task. This kind of design makes the program more flexible for using. In this dissertation, a large number of experiments have been done to test the overall system performance, a variety of comparative experimental results show that the system have higher detecting accuracy, efficiency and stability and is basically competent for inspection tasks of FPC factories.In short, to solve the problem of capturing the texture properties, several wavelet-based feature extraction methods are proposed and successfully apply them to defect detection tasks for the FPC gold surface. The study on these key technologies possesses important practical significance and brings important reference value for improving the theory and application of texture analysis techniques.