Research On Visual Attention Model Based On Significance Analysis

Visual attention is an important characteristic of human visual perception. Human can quickly grab the interesting information from seas of visual information and let the important information get processed at first priority by nervous and psychological resource. The computation model of visual attention is a method which uses computation method to imitating the working mechanism of human visual system. Currently, most visual attention models are constructed based on the analysis of visual saliency of scene. Study of computation model of visual attention is not only important to exploring the working mechanism of human visual system, but also has a great value on practical applications. As a preprocessing step, visual attention can distinguish important area from background are in an image, thus can save lots of computational resource in later image processing process. There exist two kinds of working mechanism in human visual system:bottom up and top down. The research focus of this paper is on bottom up mechanism. Existing bottom up model can be classified into spatial domain and transform domain. The spatial domain models have more biological plausibility, but suffer more from computational complexity, too. The transform domain is on the contrary. This paper first explains the research background and meaning of visual attention. Then several classical theory and computation models are introduced. And a general architecture of saliency based bottom up model summarized and the critical steps are figured out. In the following part, this paper analyzes the essence of PCT model and proposed a high frequency discrete cosine transform (HFDCT) model. The HFDCT model has very similar results with PCT model, but the computational speed is1.6times of PCT model. Besides, this paper proposes an improved PCT model:Multi-scale PCT model. The model has some improvement on feature channel selection, feature scale and feature combination. In the following part, this paper introduced the concept of local saliency and global saliency and proposes a multi-stage visual attention model. This model has good balance in local saliency and global saliency. And this model can quickly distinguish the different region or object from other regions or objects in a multi-objects circumstance.The main contributions of this paper are listed below. 1. Explain in detail the working mechanism of human visual attention. And summarized popular theory and computation model of visual attention. Propose the general architecture of saliency based bottom up computation model of visual attention and the critical steps are figured out. These will contribute much on guidance effect on later research in related fields.2. Figure out the essence of PCT model and propose a more fast model: HFDCT model.3. Present a Multi-scale PCT model which improves on feature channel selection, feature scale and feature combination. The simulation results shows the model have good performance in accordance with human visual attention behavior.4. Propose the framework of the multi-stage computation model of visual attention. This model firstly utilizes local saliency to pop out potential areas, then get area weight factor of each area to generate last combined master saliency map. This model has good balance in local saliency and global saliency. This model could label the different area quickly from an image with lots of saliency area. The idea with multistage working steps in human visual attention system has not ever shown on related academic journal, and this work has some innovating value.