A Texture Feature Extraction Method Based On Frequent Itemsets In Relative Phase Domain And Its Application In Image Classification

The multimedia technology based on images undertakes the task of transmitting information and generalizing knowledge increasingly, and it is very important to analyze and process images effectively. Texture, like other features such as shape and color, belongs to the low-level feature of an image and it can not only represent the distribution of gray levels, but also reflect both macroscopic and microscopic structures of an image. Hence it is meaningful to analyze and extract the texture feature of gray images in image retrieval and classification.Relative phase is a newly developing technology for extracting features of images in phase domain and this thesis proposes a new method of texture feature extraction based on frequent itemsets which are mined in the relative phase domain. The relative phase information can be obtained only in complex wavelet, therefore DTCWT and PDTDPB are used to decompose the original images firstly. Then the relative phases are calculated and the combination of frequent 2-itemsets features and statistical features in relative phase domain are employed as the texture features. We obtain six subbands in different directions after four-level decomposition for original images by DTCWT, and eight subbands in different directions after three-level decomposition for original image by PDTDPB. After that, we calculate the relative phase information for every subband. Then, we create the transaction database from these subbands by a sliding window and build the frequent itemset feature vectors by mining the frequent itemsets. At last, we combine the statistical features with the frequent itemset feature vectors as the texture feature to classify the images. The experiment results show that it has a good classification effect for the new method of texture feature extraction.In fact, a certain property of real images just represents a single image feature. Thus it is very practical to combine several types of image features for classification. We accomplish the image classification experiment by using Edge Direction Histogram and the texture feature proposed in the thesis together. The experiment shows that the combination of shape feature and texture feature can obtain a better result than a single feature.