As the basis of image analysis and understanding, image segmentation is one of the most underlying and most difficult problems in computer vision. Image enhancement is also very important, because its results will affect the performances of the following vision algorithms. There have been a great many of researches in image enhancement and segmentation, since the beginning of image analysis and computer vision, and there are still many papers on both topics published up to the present. In image enhancement, this dissertation analyzes two enhancement algorithms, i.e., the spatial and spatiotemporal homomorphic filter (SHF and STHF), proposed in IEEE T-PAMI in 1997 to enhance far infrared images based upon a far infrared imaging model, and proves theoretically and experimentally that the resulting images with SHF are in general smoother than those with STHF, although STHF may reduce the processing time greatly in comparison to SHE Based on this conclusion, an adaptive spatiotemporal homomorphic filter (ASTHF) is proposed. With ASTHF, the resulting images are smoother than those with STHF, while the processing time is less than that with SHF for a similar degree of convergence. ASTHF keeps the advantages of both SHF and STHF, featuring both good quality and less processing time. In image segmentation, this dissertation proposes an integrative segmentation Framework-general scheme of region competition based on scale space (GSRC). GSRC first labels pixels whose corresponding regions can be determined in large likelihood, and then fine-tunes the final regions with the help of probability model, boundary smoothing, and region competition. By means of a novel scale-space-based classification scheme, GSRC controls the extent to which an image is segmented, and establishes a quantitative relation between its parameter and the number of resulting homogeneous regions. GSRC can result in varieties of statistically homogeneous segmentation under different scales of the feature space, and also provides a formal method to group several individually statistically homogeneous patches into a single region which represents a concerned object or its background. Such segmentation is semantically homogeneous. With both semantic homogeneity and quantitative control of the number of the resulting homogeneous regions, GSRC may produce a 'clean' resulting image, therefore simplifying the following procedures. Although the description of the scheme is non-parametric in this dissertation, GSRC can also work parametrically if all non-parametric procedures in this dissertation are substituted with the parametric counterparts.
修改评论