Research On Extraction Of Roi For Image Based On Biological Visual Attention

The research of region of interest extraction is one of the hot issues in image processing and machine vision, region of interest is used not only linguistic indexing of pictures but also compression coding/object region localization and recognition/image display in small screen/image analysis of medical image. In recent years, it has gotten great progress that visual attention mechanism is applied in extracting region of interest. But it has some difference between real region of interest and the extracted region of interest. It is not exact.This paper proposed an improved method to extract more exact saliency map closed to saliency object. Study on how to express region of interest by saliency object based on saliency map. The main results of this paper can be summarized as followed:(1) We introduced the structure of human visual system and process of information. Also we analyzed two types of visual attention model: the flow chart and application of data-drive visual attention model and task-drive visual attention model. Moreover, we expounded the process of Itti's model and application.(2) We aimed at the problem that the saliency map is not exact enough by Itti's model, this paper proposed an improved method to extract saliency map more closed to saliency object. Based on analyzing the results of Itti's model, we propound a novel combination strategy among conspicuity maps; the contribution rate of each conspicuity map to the saliency map is in inversely proportional to the saliency point's area. A set of experiments were conducted over it to demonstrate the effectiveness of the proposed method. The results indicate that saliency map accuracy increased by 15%-20% compared to Itti's model.(3) Human care more about center information of image than border area. So we put forward a new optimized method to impair the saliency of border area by a location saliency map. The experimental result is good enough.(4) Based on saliency map extracted, we proposed a method to obtain region of interest. Firstly, segment image by fuzzy c-means clustering algorithm; secondly, calculate the average saliency of each segmentation area and extract the maximum saliency area as region of interest. The experiment results is satisfying, it can make up the shortage of region of interest by Itti's model.