Researches On Visual Saliency Detection Model And Its Applications

Human vision has the ability of quick search for their most interest when looking at a natural scene. This capability is called visual attention. It is believed that visual attention plays an important role in the process of human survival and development. Visual attention and how the human senses, being closely related to how we perceive and process visual stimuli, is investigated by multiple disciplines including cognitive psychology, neurobiology, and computer vision. With the continuous development of cognitive psychology and neurobiology, based on the research of the mechanism of vision, human vision on the selectivity of the target in the scene can be divided into two manners:A rapid, bottom-up, saliency-driven, task-independent manner and a slower, top-down, volition-controlled, task-dependent manner. Visual saliency is closely related to visual attention, which is a key step of visual attention mechanism. Salient region detection can be described as quickly locating salient regional and measuring the saliency of salient regional. Salient region detection is widely used in many applications of image processing including image segmentation, object recognition, adaptive compression of images, content-aware image editing, image retrieval, object detection, object tracking and image quality assessment. In this thesis, the key technologies of visual saliency detection and its application are studied from visual attention mechanism. At the same time, we propose some new ideas and algorithms of visual saliency detection. The main work and contributions of this thesis are listed as below:1. In view of the disadvantage of the existing global features and local features, this thesis proposes a salient region detection algorithm by integrating global features. Firstly, our method uses the uniqueness and the color spatial distribution global features to compute the corresponding saliency maps. Secondly, fusion in the framework of conditional random field with a set of saliency maps, through salient region and the background labeling achieve initial salient region detection. Then, we compute object prior map with the gaussian model based on salient region, and filter the global feature of saliency maps with the gaussian filtering. Finally, filtered saliency maps are fused by the conditional random field for achieving more accurate detection. Experimental results show that the proposed approach can generate more accurate saliency maps with uniformly highlighted foreground and well suppressed background, and effectively improve precision and recall.2. Excavate the high level saliency prior features based on visual mechanism. This thesis proposes a novel salient object detection algorithm by integrating multi-level features including local contrast, global contrast, and background priors which measure the visual saliency in pixel-level, region-level, and object-level. We use the low level visual cues based on the convex hull to separate salient object from the background. The background priors are computed from the background templates using Principal Component Analysis. In order to suppress background noise, local and global contrasts are refined by object center priors which are computed with the Gaussian model based on background priors. Experimental results on widely used public benchmark datasets demonstrate the proposed method can better highlight salient object when against state-of-the-art methods. We also demonstrate Otsu adaptive threshold method can be used to create high quality segmentation masks.3. In view of the application of visual saliency in the process of object tracking, this thesis proposes an object tracking algorithm based on visual attention. Firstly, the proposed algorithm uses visual saliency detection to extract the saliency features. Secondly the proposed algorithm extracts the object’s motion features based on bayesican decision theory of foreground and background classification method. Then, we use color features and motion features to estimate the target state under the guidance of saliency features. Finally an adaptive particle filter is used to fuse these features for object tracking. Numerous experiments demonstrate the proposed method in object tracking under complex scene has stronger robustness, and when dealing with illumination change, pose variation, occlusions, rapid movement, and complex scene situations, it has good tracking effect.4. In view of the initial location problem of the relay tracking under linkage of box camera and dome camera when the target is leaving the scene of the box camer, this thesis proposes a relay tracking method under linkage of box camera and dome camera based on visual attention. The method uses interpolation algorithm on grid to zoom in the tracking target under the dome camer. When the target is leaving the scene of the box camer, visual saliency detection algorithm is used to calculate the candidate regions. We use the saved template of target under the scene of the box camer to match the target in the candidate area, and get the position of target under the dome camer. Finally, active tracking is realized by using Mean Shift tracking algorithm. The experimental results show that the proposed relay tracking under linkage of box camera and dome camera based on visual attention has a good real-time tracking effect.