Image Segmentation Method Based On The Theory Of Multi-scale Study

Image segmentation plays a very important role in image technology. Because of its importance, it is always the focus of the image research. Although a lot of researching work has been done on it, the breakthrough has not shown up. Difficult problems of Image segmentation exist in the dilemmas between the segmentation accuracy and the segmentation easiness, and between the over-segmentation and the under-segmentation. These problems are not easy to be solved by segmentation technology in single scale. A large number of phenomena or processes possess features at multiple scales in nature or projects, so multiscale analysis method can capture the several important characteristics of phenomena or processes in essence. Through integrating with image information in different scale, multiscale analysis method should unify the segmentation accuracy in finer scale and the segmentation easiness in coarser scale, in our opinion, this analysis method is suitable for automatic or semi-automatic image segmentation.In this dissertation, we attempt to have an in-depth investigation on the application of multiscale analysis method in image segmentation and develop the research work of the following three aspects. Firstly, from the angle of scale space, the image pre-processing based on multiscale method is realized by using wavelet transform, which lays a foundation for the consequent image segmentation. Secondly, multiscale autoregressive model is introduced into image segmentation and combined with wavelet transform to establish the mathematical relations between levels and between pixels in neighboring levels in order to construct a more effective multiscale data structure. Thirdly, we explore the multiscale segmenting tactic. The proposed approach uses multiscale MRF model in image segmentation stage, which improves the accuracy and efficiency of image segmentation. Some improvements on multiscale MRF model are proposed based on existed work, which make the model more suitable for multiscale image segmentation. The main research and contributions of this dissertation include:1. A morphological scale space filtering approach based on wavelet transform is proposed. It builds up a morphological scale space on the basis of structure elements of different scales. Then wavelet transform is utilized to integrate the information in different scale and eliminate the drawback of morphological scale space method. This filtering approach not only suppresses noise but also preserves the boundary of the target well. Then on this basis, we propose a fast watershed segmentation algorithm and solve the over-segmentation problem of traditional watershed method.2. With analysis on the dynamic system model based on wavelet transform, the multiscale Kalman filtering approach based on wavelet transform is proposed. It uses the multiscale characteristic of wavelet transform to decompose the initial predicting sequence and perform Kalman filtering estimate on each level. Then wavelet reconstruction is applied to integrate the estimate information from each level to obtain a more accurate estimate. The proposed method combined the wavelet transform with the layered Kalman filtering and thus achieves a better state estimate filtering result.3. After research on multiscale autoregressive (MAR) model, we analyze the advantages of constructing multiscale data structure with this model and then propose a SVM segmentation algorithm for SAR images based on MAR model. According to the characteristics of SAR imagery and the self-similarity of targets in SAR images in different scale, establish MAR model of the SAR images and extract multiscale feature vectors of the images through MAR model. Finally, a generalized weighted SVM classifying method is presented to perform classification for multiscale feature vectors.4. In order to solve the two major problems of the complexity of parameter estimate in using the distribution of conditional probability and the difficulty of deducing an accurate distribution in theoretical way, the segmentation approach based on MRF and SVM posteriori probability is proposed. In terms of Bayes formula, it converts the estimate for conditional probability into posterior probability estimate. After training by samples, it generates posterior probability by mapping the output of SVM decision-function and brings the information of posterior estimate by SVM into MRF model to perform image segmentation.5. A novel multiscale MRF image segmentation method based on context information is proposed in order to solve the problems of some little classification mistakes in uniform region and the difficulty of segmentation accuracy of region boundary. Through estimating the context parameter approximately, the algorithm realizes the mutual information of the same neighboring nodes and adjoining upper scale neighboring nodes. The advantage of little complexity added to the model is important. Segmentation results are greatly improved both on segmentation accuracy and boundary localization for context information.6. By means of combining MAR model with multiscale MRF model, the dissertation proposes a new semi-tower model to segment the images on the basis of the above two models. In image segmentation, MAR model is introduced into multiscale MRF model, and then the double semi-tower structure of label fields and observation fields is established, this method we presented improves the accuracy and efficiency of image segmentation.