Research Of Robust Image Retrieval Algorithm Based On Multi-Features

With the dramatic development of multimedia and internet techniques, image data has covered the every inch of our daily life. How to search the interesting and useful information from the huge image data collections has been a delicate problem. Content-Based Image Retrieval (CBIR) technique plays an important role to cope with this problem.One of the keys in CBIR is the feature extraction, but most methods for feature extraction ignore the robustness to noise and geometrical distortion, this thesis focuses on the robustness of multi-feature extraction, and the main research works are summarized as follows:1. A robust color feature extraction algorithm is proposed. According to the image bit-plane theory, the impact of the noise attacks focuses on the lower bit-plane. Firstly, the primary bit-plane image is extracted from the original color image. Secondly, the color histogram is calculated from the primary bit-plane image as color feature for retrieval. Experimental results show that the proposed method is robust to the many noise attacks including blur, sharpen, lighting, etc.2. A robust texture and edge features extraction algorithm is proposed. First, the noise-mitigation bit-plane image is extracted from the Y component of YCbCr color space to mitigate the impact of noise on edge feature and maintain the texture and edge characters to a large extent. Then, a Robust Object Region (ROR) is determined in accordance with the image normalization theory in order to be invariant to geometric distortion. Finally, a wavelet-based texture feature and an edge histogram are extracted from the ROR respectively. Experimental results show that the proposed method is robust to many geometric attacks including translation, rotation, and scaling, etc.3. Combining the above two approaches, a robust multi-features extraction algorithm for image retrieval is presented. An empirical study shows that the proposed algorithm is not only effective to retrieve target images the user want, but also robust to noise and geometric attacks.