Research On Salient Object Detection Based On Regional Feature Integration

With the rapid development of information technology, the image has becomeone of the most important information, the scale and complexity of image data areincreasing. Facing such huge image data in big data era, how to effectively deal withimage information has become a hot topic in the field of image processing. Humanvisual system (HVS) has a remarkable ability to automatically pay more attention toimportant visual stimuli in natural complex scenes, This ability for prioritizingimportant information, also known as visual attention mechanism, makes humanfulfil various visual tasks quickly, precisely and efficiently. In order to find out ananalogous model mimicking human attention mechanism, researchers in physiology,psychology and computer vision have been making a good effort for a long time andproposed many computational models for saliency. Salient object detection technoloybased on visual attention mechanism plays a significant role in several fields, such asimage procesing, artificial intelligence, compuer vision and so on.This thesis firstly describes the remarkable significance of salient objectdetection, then analyzes the research status in this field and describes the biologicalmechanism of visual attention and visual attention modeling theoretical foundation.Secondly, introducing several kinds of visual features of salient objection,including color, depth, texture and shape features, as well as methods of extractingthese features. Microsoft Kinect and RGBD salient object detection benchmark aresummarized in detail.Thirdly, proposing RGBD salient object detection method based on regionalfeature integration. This method adds regional depth information of image to saliencycomputation, to apply to object detection. Firstly, we get several region throughmulti-level image segmentation. Then, we build the regression random forest bylearning varieties of regional features, and use the supervised learning approach tomap the regional feature vector to a saliency score. Finally, we fuse the saliencyscores across multiple levels by least square method, yielding the saliency map. Experiments show our method can accurately locate the salient objects from RGBDimages.Finally, the full text is summarized and prospected.