Research On Some Approaches To Semantic Object Segmentation

The application domain of semantic object segmentation is quite broad. It includes video coding, video surveillance, image and video editing, image and video retrieval, human computer interaction, image understanding and object recognition, etc. The performance of these application systems depends largely on the results of semantic object segmentation. Therefore, the research on semantic object segmentation is very important for a variety of applications. On the basis of traditional video object segmentation, the research on semantic object segmentation has gradually been focused on automatic extraction of salient objects and visual attention based interesting semantic object segmentation in the recent years. Although many progresses have been achieved in this research field, there is still room for improvement. The research in this dissertation mainly belongs to this research field and has improved some related approaches. In addition, moving cast shadow has also been studied in this dissertation, and good results have been obtained. The main work of this dissertation is as follows:(1) An efficient salient object extraction approach based on region saliency ratio is proposed to automatically segment visually salient objects from color images. The input image is first segmented into homogenous regions using nonparametric kernel density estimation, meanwhile a scale-invariant saliency map is constructed based on multi-resolution feature contrast calculation. Then the region saliency ratio of each region combination to its complement is calculated in turn. Finally, salient objects are extracted by maximizing the region saliency ratio. The proposed approach can efficiently extract multiple salient objects complying with human vision attention from color images.(2) An segmentation approach that combines selective visual attention with the Markov random field (MRF) framework is proposed to segment interested moving objects from video sequences. The input images with 256 gray-levels are first transformed into the images with 8 gray-level bands, and motion feature is extracted based on variation of gray-level band between two consecutive frames. Then moving regions are obtained by combining motion features with connected component labeling. The shape features of moving regions are extracted and compared with the predefined shape features of interesting objects, and initial object mask is generated by a set of selected moving regions. Finally, the memorization along time for each pixel is used as the energy function of MRF and more accurate moving object segmentation results are obtained by energy minimization. The proposed approach can efficiently distinguish moving object pixels from background pixels, and improve the accuracy of object segmentation.(3) A moving cast shadow removal approach based on chromaticity, intensity and the edge information is proposed for accurate video object segmentation. Based on kernel density estimation and edge information of the input frame, an initial moving object mask and corresponding edges of moving objects are obtained. Then candidate shadow regions are obtained by extracting chromaticity and intensity information from the input frame. Finally, the moving cast shadow region is detected and removed using region growing method. Experimental results of indoor and outdoor video sequences demonstrate the good performance of moving cast shadow removal of the proposed approach.