Research On Region Of Interest Extraction

Human vision system often focuses visual attention on some special objects of the scene when processing a relative complex scene and processes these special objects in priority so as to get the main information of the scene in minimum time. This process is called vision attention. The region of which the special objects in the scene compose is called region of interest(ROI). With extracting and analyzing ROI of images, we can effectively reduce calculation amount of image process and improve the efficiency of information process. ROI extraction has important application on image compression and coding, image retrieval, target detection and recognition, scene analysis and active vision, and so on.This thesis mainly focuses on the problem of ROI extraction and makes further research from eye movement experiment data, low-level image feature analysis, vision attention model and high-level semantic. The research work in this thesis consists of four parts as follows:(1)This thesis firstly proposes an ROI extraction algorithm based on eye movement data. With designing eye movement experiment and using eye tracker, we obtain eye movement data of testees. With data selection and coordinate transformation methods we get effective fixation data furtherly. Then with of Gaussian distribution simulation, focus of attention searching with "marking helms" method, seed filling, the eye movement ROI of testees is extracted. Eye movement ROI is selected as the real ROI of image and is also applied on effectiveness evaluation of other ROI extraction algorithm.(2) Based on color feature analysis, this thesis researches the influence of multiple low-level features combining with vision attention to ROI extraction and proposes two ROI extraction algorithms based on color feature and best weights of low-level feature respectively. The algorithm based on color feature designs eight color feature with color theory and combines vision attention model to get computing saliency map. Then it compares the computing saliency map with saliency map generated from eye movement data. With the comparison the optimized color feature selection rules and image segmentation algorithm are designed to implement automatic and accurate ROI extraction. The algorithm based on best weights of low-level feature aims at color, brightness, direction and texture low-level features of images, makes regional evaluation between ROI extraction from eye movement and low-level features respectively. Then with analyzing the influence of four different low-features to ROI extraction of different images and the best weight of low-level features in ROI extraction is got.(3) With effective analogy between eye movement experiment and eye trajectory, this thesis proposes two ROI extraction algorithm based on salient points and salient region respectively. The algorithm based on salient points adopts statistical method and K-means cluster algorithm, detects image contour with morphological methods and selects cluster center as seed points to make seed-filling operation to outline drawing. Then the outline drawing is made mask operation with the input image to implement ROI extraction. The algorithm based on salient maps improves Grabcut image segmentation algorithm and uses salient maps generated with vision attention model to replace the user's interaction instruction to the foreground and background of image and implement ROI extraction.(4)This thesis proposed a ROI extraction algorithm with combination of bottom-up and top-down visual attention (VA). The proposed algorithm transforms the top-down information such as observation task et al. to the weights of low-level features. Then with Itti and Stentiford bottom-up vision attention model to implement ROI extraction which combines observation task and vision stimulation of images. With transform the query attention to the weight of saliency map, i.e. with setting the corresponding weight of low-level features to influence the solution of saliency map, the gap between observer's query attention and low-level features of images is effectively shorted and the extracted ROI accord with observer's requirement furtherly.