Research On Salient Region Detection Algorithm

Human beings are quite adept at swiftly detecting objects of interest in complex visual scenes with the assist of Human Visual System (HSV). Simulating human visual system to detect visually salient regions of an image, which is also known as salient region detection, has been one of the active topics in computer vision. In essence, the key issue related to salient region detection is the deployment of visual attention model. In light of the massive studies in the past years in neuropsychology, the deployment of visual attention has long been believed that there are two different approaches in visual processing mechanism:bottom-up approach and top-down approach. Different from the top-down approach, which requires some prior knowledge on the task to be fulfilled, the bottom-up approach is data-driven and task independent.The work of this paper mainly focuses on the bottom-up visual attention model for salient region detection and some research results include:1. For complex natural scenes, time domain and frequency domain based saliency detection algorithms will generally lead to poor integrity of the detected region and blurred edge, respectively. By taking advantage of the local contrast cue from time domain and the global frequency spectrum cue from frequency domain, we propose a time-frequency domain integrated saliency detection algorithm. Thus, the discrimination between the salient region and the background can be effectively enhanced.2. By posing the salient region detection as a novelty detection problem, this paper proposes an ensemble dictionary learning based salient region detection framework. Under this framework, a number of random over-complete dictionaries are trained independently, obtaining multiple sparse representations with good adaptability on the image itself. To boost the contrast between the salient object and the background area, a novel reconstruction error based dictionary atom reduction method is presented with low computational complexity. Meanwhile, we also provide a good interpretation on the proposed framework from the perspective of Bayesian probabilistic theory.3. This paper presents a video salient moving target detection method by considering the spatial-temporal consistency. Specifically, the static saliency detection based on spatial information is first carried out to predict the moving object in a video frame, and then the statistical entropy of adjacent video frames is employed to make further verification. Compared with frame subtraction method for moving target detection, the generation of "hole" in a detected salient region can be effectively avoided.