Rotation, shift and scale invariant wavelet features for content-based image retrieval and classification

Rapid continual advances in computer and network technologies coupled with the availability of relatively cheap high-volume data storage devices have effected the production of thousands of digital images or photos everyday, e.g. in World Wide Web. Many content based image retrieval (CBIR) systems have been proposed to cope with such huge image archives. However, most of the current content based image features for CBIR do not address the issues of orientation, position or scale variations of images. In practice, the presence of orientation, position or scale variations of images in digital image repositories could introduce problems in image retrieval. The neglect of such problems could result in a relatively poor performance of the current CBIR systems.; Discrete wavelet transforms have been shown to be effective for image analysis. But it is well known that one major drawback of discrete wavelet transforms is their lack of rotation, shift and scale invariance, and joint invariance. The work presented in this thesis focus on the extraction of an effective and robust rotation, shift, scale and joint invariant wavelet features for content-based image retrieval and classification. An adaptive rotation invariant wavelet packet transform and seven schemes for extracting the invariant wavelet features have been proposed. The invariant wavelet features are represented by the dominant energy signatures computed by the respective invariant wavelet coefficients generated. The feature extraction process is quite efficient, with an overall computational complexity of O(n · log n), where n is the number of pixels in the image. Experimental results for CBIR show that the proposed invariant wavelet feature can achieve a high retrieval precision and recall for general performance testing and can achieve perfect retrieval precision for invariance performance testing. Moreover, experimental results for content based image classification (CBIC) show that the proposed invariant wavelet features can achieve very high classification accuracy for both general performance testing and invariance performance testing. The results also show that the proposed invariant wavelet features are quite robust to noise and outperform other image classification methods significantly.