Image Recognition Based On Wavelet Transforms And Fuzzy Rough Sets Technique

Image recognition is the process of image cognition and image understanding by using modern information processing technology and computer technology, which has attracted lots of attention in pattern recognition and image processing during the last two decades. Image recognition mainly investigates the image classification based on image features and has been widely applied to many fields. It roughly consists of three steps: image preprocessing, image feature extraction, and image classification. The image feature extraction and image classification are two vital steps, which are the major focuses of this thesis. The contributions of this thesis in image feature extraction mainly include the following two aspects:(1) This thesis integrates the wavelet transforms (WT) and two-dimensional projection subspace techniques effectively. The selection of wavelet sub-band images used for image recognition is studied. Three methods for image feature extraction are proposed,â‘ image features extraction method based on WT and two-dimensional principal component analysis (2DPCA),â‘¡image features extraction method based on WT and two-directional two-dimensional principal component analysis ((2D)2PCA), andâ‘¢image features extraction method based on WT and two-dimensional linear discriminant analysis (2DLDA). The experimental results on four face databases (ORL, YALE, JAFFE, and UMIST) verify the effectiveness of the proposed methods.(2) As a matrix decomposition method, singular value decomposition (SVD) can be used to extract algebraic features from images. The SV features of images have many good properties such as stability, geometric invariance, and insensitiveness to noise. However, it is difficult to achieve high recognition rate by only using one scale SV features in image recognition. Based on the wavelet transforms and SVD, this thesis proposes an image feature extraction method which combines multiple scale SV features of wavelet sub-band images. The recognition rates on three face databases (ORL, YALE, and JAFFE) are 82.11%, 100%, and 95.68% respectively, which are higher than the existing SVD based approaches.For image classification, this thesis proposes three classification methods based on fuzzy rough sets technique.(1) To handle the redundant features which are extracted from images by using 2DPCA methods, an image recognition method (image classification method) is presented based on 2DPCA and fuzzy rough sets technique. The proposed method selects the important features for classification by using attribute reduction in fuzzy rough sets theory. The experimental results show the proposed method outperforms the existing image recognition methods based on 2DPCA.(2) Fuzzy rough sets are generalizations of rough sets to deal with both fuzziness and vagueness in data, which integrate fuzzy sets and rough sets together. Based on fuzzy rough sets technique, this thesis proposes a new criterion, in which expanded attributes are selected by using significance of fuzzy conditional attributes with respect to fuzzy decision attributes. Because that two uncertainty (fuzziness and roughness) are integrated together in the proposed method whose power of classification is higher than the one of fuzzy decision tree algorithms based on pure fuzzy entropy. An illustrative example as well as the experimental results statistically confirms that the proposed method is superior to the fuzzy ID3 algorithm.(3) Given a fuzzy information system, we may find many fuzzy attribute reducts and each of them can have different contributions to decision-making. If only one of the fuzzy attribute reducts, even though the most important one is selected to induce decision rules, some useful information hidden in the other reducts for the decision-making will be losing unavoidably. To make good use of the information provided by every individual fuzzy attribute reduct in the fuzzy information system, this thesis presents a novel induction of multiple fuzzy decision trees based on fuzzy rough sets technique. An illustrative example as well as the experimental results validate that the proposed multiple tree induction has better performance than the single tree induction based on the individual reducts.In addition, this thesis investigates the influences of the coarse granularity and the fine granularity on the decision tree induction. The investigation leads us to the conclusion that the information entropy under coarse granularity is not less than the one under fine granularity. Furthermore, we draw the conclusion that the decision tree generated by selecting the expanded attribute under fine granularity outperforms the one under coarse granularity.