The Study On Object Detection Based On Visual Saliency

The visual system will autonomously detect and perceive sensitive information from every position of the scene, whlie we observe external environment, which is usually called visual saliency. The principles is derived from the bionic research of the human visual system, then researchers construct computation model based on the principles of cognitive neural, simulate the process of capturing and dealing external information, and achieve salient object perceiving. The versatility and efficiency make the saliency visual computing one of the most important research branch in the object detection field. This thesis analyzes and summarizes the existing work in saliency research, and improves deficiencies of existing models in complex scenes, to make better detection effect.Simple target saliency detection has achieved good results, but when the scenes become complex, single image saliency detect could not make accurately object area detection and judgment, combining and using multiple related images can fuse more information, significantly suppress complex background, highlighting the salient objects, and improving object detection quality. Video sequences as the most common images set in daily life, the motion information in video is very important perceive clues in human visual system, so making use of the motion characteristics saliency on research of video salience detection is a key point of this thesis. Meanwhile, different perspectives of a scene can also provide compensation information in the target and background area, this thesis will also research on multi-view image saliency.The main contributions are summarized in the following two points:1. Video sequences saliency object detectionUsing correlation in two time continuous frames, to extract motion feature,analyzing the motion amplitude and motion direction of each position respectively, fusing motion features and static color contrast features, construct a spatial-temporal saliency object detection method, highlight the visual system sensitivity to the motion of target. Compared to the existing models, this method has better detection results in a number of test video sequences.2. Multi-view image saliency object detectionExtending single visual view to multi-view in a scene, analyzing the spatial relationship between the foreground and background in different perspectives, to get the key argument that the surrounding background of the target in different perspectives has different occlusion regions, and applying it to saliency object detection, experimental results show that this multi-view combined saliency object detection method can get clearly object boundaries, and effectively suppress the complex background around saliect object.