Object Salience Detection Based On Visual Attention Mechanism

Object salience detection is an important problem in Vision. It is a basic research way to detect salient regions which can imitate the human-eye vision characteristic. Although there exist a number of working algorithms and systems for static image salience, to develop a salient region detection algorithm for video, is still challenging. One important feature of human vision mechanism is that for video human not only tends to pay attention to the static saliency including brightness, color, contrast to surrounding areas, but also pay attention to the motion saliency of the object.Vision Research is interdisciplinary, covering a variety of fields from Computer Vision to Neurobiology. This thesis is about research work on visual attention features and object salience detection algorithms with an emphasis in the regime of Computer Vision. We firstly make a survey on recent achievements of salience detection. In addition, one primary approach, the bottom-to-up model, is further studied. Then the detection of the motion salience in video and the static salience in image is elaborated and analyzed in this thesis. Based on the insights, an object salience detection algorithm is developed and experiment results show that it is indeed practical.The main contributions of this thesis are as follows:1. Research and compare the advantages and disadvantages of object salience detection models and algorithms.2. Based on the feature of visual attention, we propose an algorithm to detect motion salience in video. We firstly partition the video to get a series of frame segments, then detect the salience feature from motion vectors by block-matching algorithm. A motion salience detection algorithm is then given based on the motion vector.3. Practical models and solutions of static image salience detection are given. We analyze the spectral residual and incremental coding length algorithms. Compared to traditional feature-based methods, they show more distinct salience regions and relation with neurobiology.4. The integral object salience detection algorithm is designed based on the visual attention mechanism. We propose a fusion rule to construct the integral salience according to the visual attention effects of motion salience and static salience. The experiment results prove that this algorithm leads to good performance and thus is applicable.