The Research On Class-Speciifc Semantic Image Segmentation

With the fast development of multimedia and internet technologies, semantic web willbe the next stage of the evolution of the present network. Semantic image segmentation isthe basis to build semantic web, especially to obtain the content-based representation ofvideos and images. Our research is mainly used on cow recognition system for videoprocessing. The important step is to operate the cow class-specific image semanticsegmentation. Recently, there are still many big challenges on how to map low-levelfeatures to high-level semantic features, so-called “semantic gap” problem. In this thesis,we use the “Bag-of-Words” model and combine both merits of image-based andclass-based segmentation algorithms. The proposed algorithm can find reasonable andefficient mapping model to realize semantic segmentation of cow images.Based on the existing research technologies, we mainly concentrate on the followingwork:1) We propose an improved image segmentation algorithm based on the watershedtransformation and multi-resolution analysis. We firstly discriminate the real regionmaximum and construct the ideal marker model by utilizing morphological reconstruction;and then complete domain fusion by the third-moment of wavelet coefficients; finally, weuse the inverse wavelet transformation to project the low-resolution image intohigh-resolution.2) Based on the Top-Down segmentation (CSF-SEG─Class-specific FragmentSegmentation), we present a novel approach to extract fragments. We simultaneously makeuse of the CSF training and mapping libraries to guide the new image segmentation, andsearch the optimal cover for the image based similarity criterion, thus realizing the imagesegmentation.3) Combining the merits of Top-Down and Bottom-Up approach, a novel andefficient segmentation method based on “Bag-of-Words” is presented. We firstly use the improved Top-Down algorithm, which is based on Bag-of-Words model, to fulfill theinitial segmentation. In this procedure, we employ the wavelet-based homomorphicfiltering algorithm to remove the illumination effect for the training and segmented images.Secondly, we take the result of the initial segmentation as the input of Bottom-Upsegmentation, and use the improved multi-scaled marker-controlled watershedtransformation to refine the boundaries. The novelties of the proposed algorithm aregenerally given as follows: Firstly, we propose to capture SIFT descriptors and form avisual vocabulary after K-means clustering based on “Bag-of-Words” model. Also, byusing of similarity criteria, we extract effective visual words as the object-based elementCSF, and map them to the original images. In the second, we utilize the improvedimposition minima technology to build an ideal marker model.