Salient Object Detection Base On Visual Salient Feature

Visual saliency is can be represent salient information of scene images，which is acquiredby extracting visual salient feature and using cognitive mechanism. The research of visualsalieny is base on visual cognition of human; it is can be humanly represent information ofimages. Saliency map is based on human visual system (HVS), and provides guidance to usfor selecting a few salient visual regions quickly to pay more attention to useful information.Visual saliency model can assign the limited processing resource in system priority to a fewsalient visual regions. The research of visual saliency computational model is proposed basedon analysis of human visual behavior，that is how to perceive the world by their own visualsystem。 According to physiology, psychology, and eye movement tracking method，weacquire a large number knowledge about human visual perception: Which feature informationstimulates eyes of human more strongly. How to process low level visual and construct highlevel information. Acquire high level information by which cognitive mechanism. Visualcognitive behavior include which general rules.How to use these general rules andmechanism process low-level feature,and acquire high level visual information. With thedevelopment of HSV，the knowledge of visual cognition is Constitute a complete system.Visual stimulation principle from feature information of image，visual cognitive mechanismfrom brain of human and general visual behavior rules,all of them can construct visualsaliency model，and use for object detection.The research is beginning to visual salient feature, select some corresponding feature ofimage based on cognitive physiology and cognitive psychology. According to research ofvisual cognitive mechanism and behavior rule，proposed visual saliency model for objectdetection. In this research, we analyze the feature of dynamic object and static object，construct model for detecting dynamic or static object. The model can be used in naturalscene and Visual robot platform.In research of static object detection, for detecting object from complicated background,and rapidly segment salient object from scene image, proposed an adaptive segmentation forsalient object detection. It is inspired by an assumption that image information consist of redundant information and novelty fluctuations， in addition object detection is to remove thenon-saliency part and focus on the salient object. Considering the relation betweencomposition of image and aim of object detection, we construct a more reliable saliency mapfor evaluating composition of image. Local energy feature was combined with simplebiologically-inspired model (color, intensity, orientation) to strength the integrity of object inthe saliency map. Estimate entropy of object via maximum entropy method. According to thesaliency map, remove pixels of minimal intensity from original image. Then compute entropyof the images removed redundant and correlate with the entropy of object for final result.Experimental results show that proposed algorithm outperforms the state-of-the-art methods.We compute Recall Rate and precision rate, and acquire ROC curve and F-Measure forevaluate method.. Experimental results show that proposed algorithm outperforms thestate-of-the-art methods.The objects in natural scene images are irregular contour. For segmenting object frombackground more accurately, we propose a salient object segmentation method based onlow-level visual feature and middle-level visual cues. Extract the low-level visual feature oforiginal image via color, intensity, orientation and local energy feature channels, combine fourchannels for the saliency map. Acquire salient feature mask via maximization informationentropy principle. In mid-level visual cues, we use an over-segmentation of an image intosuperpixels. According to salient intensity automatically set initial parameters and consideringspatial information of feature vector in clustering, it can be more accurate support contour ofobject. Finally，for segmenting irregular object from background,, classify superpixels byusing salient feature mask, that means low-level and mid-level visual feature fusion. Theexperimental results demonstrate that the proposed method is less sensitive to complexillumination and background, and can segment contour accurately. Moreover, It can be used tosegment Irregular objects from the complex background.In research of dymanic object detection, we propose an opponent color optical flowalgorithm for robust object tracking. Unlike previous works of coloroptical flow methods thattreat color as separating channels, the proposed algorithm exploits opponent colorrepresentation of color and processes color as a holistic signal. In this way, it enables moreaccurate flow estimation at the pixel locations of spatial color variations, and reduces trackingerrors by leaving more features points at their correct locations on the objiect. For successfuland efficient object tracking, we also proposed a novel type of opponent color corners that arereliable features during tracking. Together with grayscale corners, they form a good feature point set, especially when used with the proposed opponent color optical flow algorithm. Weconduct a quantitative evaluation on publicly available dataset to verify the efficacy of theproposed algorithm. And object tracking experiments demonstrate that robust tracking can beachieved for dymanic object detection.Use the adaptive segmentation for salient object in visual robot platform.because it ismore difficult to create the most advanced detection, obstacle locating and detection methodsfalter for different environment. The joint use of multiple imaging modalities is one means ofimproving some of the quality measures of the input data, in hopes of ultimately improvingoverall system performance. Since integrating radar with a Vision Sensor for obstacledetection will presumably improve overall detection performance. Combining them can beadvantageous, especially when one algorithm performs poorly in detecting objects. Forexample, salient analysis has an obvious limitation of the image in which the objects are noobstacles，and fails in detecting any salient area as object. To deal with the above challenge,the proposed method uses salient detection model for locating and combination ofpre-location mask from infrared sensors information about the objects. For extracting integrityfeature of object, we make local energy feature instead of orientation feature. That proposedcan be helpful for obstacles location. According to the issue of salient detection, we useinfrared sensors for detecting whether objects are barrier. The pre-location mask from theresult of infrared sensor is useful for efficiency of algorithm. The object location method isbased on object segmentation. The fusion of radar-vision is as pretreatment, and then wedesign a maximum entropy estimate method for obstacle location. Finally, we test our radar-vision system in indoor simulation environment.