The Research Of Image Retrieval Methods Based On Muti-resolution And Saliency Feature

With the development of the science technology and human civilization, the visual sense of people is not only limited to the information based on text, also need these information from three-dimension even higher dimension technique. The innovation of digital technology ensured the store and transmission for a mass amount of data from pictures or video, which strongly facilitates people daily life. However, how to manage efficiently and lookup accurately this information is the critical issue for image retrieval technique to be resolved. Nowadays, the image retrieval technique mainly focused on content-based. There are several popular issues in this approach: based on compressed domain, based on visual attention, based on semantic modeling, based on fusion of multi-modal features, based on relevance feedback techniques. We mainly try to develop the following research work based on multi-resolution in the compressed domain and saliency feature in visual attention.Firstly, the image retrieval method based on multi-resolution color and shape feature is presented(MRCS). The proposed method combines wavelet transform and coherence vector technique. The dynamic.threshold is applied to replace the fixed threshold in the coherence vector in order to determine the pixels either coherent or incoherent, and the de-nosing method by soft threshold and the Otsu method are also used to optimize the subband wavelet coefficients. The revised circularity is adopted to extract the geometric feature of image in order to improve the high sensibility of the change of shape. Experiment results obtained by the method on Corel Database show that the total average precision of 72.5% and the total average recall of 42.9% are attained.Secondly, the image retrieval method based on multi-resolution and saliency feature of the point is presented(MRCSP). The Harris corner detecting algorithm is improved by selecting appropriate similarity value to reduce the running time and keep the accuracy. At the same time, the wavelet saliency value is improved by fusing wavelet coefficients of the high frequency band in different directions. Image feature in the frequency domain is extracted by the method mentioned above and then is applied to image retrieval. Experiment results obtained by the method on Corel Database show that the total average precision of 80.8% and the total average recall of 47.6% are attained.Thirdly, the image retrieval method based on multi-resolution and saliency feature of the region is presented(MRCSR). The residual between the original image and reconstructing image by the orthogonal matching pursuit algorithm is extracted by Fourier transformation. The edge feature of image could be expressed by the high frequency band. The saliency feature of image is described by fusing the residual and edge feature. The revised circularity is used to extract the region feature of image and then is applied to image retrieval. The coherence vector improved by the wavelet transform is still adopted in the low frequency band. Experiment results obtained by the method on Corel Database show that the total average precision of 83.1% and the total average recall of 50.9% are attained.The image retrieval method based on multi-resolution color and shape feature is presented in the frequency domain. And then, the image retrieval method based on multi-resolution and saliency feature is also presented by combining the visual attention model. A series of image retrieval method are established in frequency domain. The results in the thesis can provide new ideas and feasible method for the researcher as well as provide a reference for the application of image retrieval technique.