Image Retrieval Based On Wavelet Multi-resolution Analysis

Image retrieval based on wavelet multi-resolution analysis is one of the active fields in content-based image retrieval recently. In this dissertation, our research focus on how to derive features from the multi-resolution representation of integer-to-integer wavelet effectively and compare different affects of retrieval performance which were brought on by different similarity measures.Firstly, this dissertation introduces low-level features extraction of an image, similarity measures between two images, relevance feedback technique and so on. Secondly, the basic principles of wavelet transform are studied, in which the integer-to-integer wavelet and wavelet packet are emphasized especially. At last, some exploratory research work has been done from the two aspects mentioned above. The main contributions of this dissertation are summarized as follows:Firstly, an image retrieval approach based on color and texture features is proposed. Based on the special disposal on the HSV color space, the color histogram taken from low frequency band is used as color feature, the statistics information of high frequency band is used as texture feature. Experiments with different kinds of color flower image database indicate that the synthesized features are superior to the single feature in image retrieval, the proposed algorithm takes full advantage of the rich representation of color, the multi-resolution and the statistics of the wavelet coefficients.Secondly, we proposed a multi-feature image relevance feedback retrieval algorithm based on color and texture features. The circular region F-norms in low frequency band of wavelet transform is used as color feature of the image and the energy ratio of high frequency band is used to form texture descriptor. Furthermore, relevance feedback technique is used to adjust the weight of similarity measure gradually in order to satisfy the user's retrieval goal. Experiments done in color image database including banners, butterflies, grassplots and so on suggest that the proposed method improves retrieval performance, with low-dimensional features and few computation merits. It is very suitable for large image database retrieval.Thirdly, images in medical database are gray, so color feature often used has no effect on medical image retrieval. Texture structure of body's various organs are very much different, whereas which of the same one is quite distinct. Hereby, an image retrieval algorithm with texture feature taken from different sub-bands of the integer-to-integer wavelet transform is putted forward, which derived satisfactory retrieval results.Next, for texture image database, traditional pyramid wavelet transform is inadequate for describing multi-resolution and multi-direction characteristic of texture image. We come up with a texture feature retrieval method based on integer-to-integer wavelet package transformation. This algorithm takes full advantage of the intrinsic...