Computational Model Of Visual Attention Mechanism And Its Application Research

In recent years, how to mimic human visual system to do research on machine visual attention model has aroused wide concern in image processing domain and related fields. This paper, based on relevant biographic materials, mainly deals with the analysis of different visual attention models and particularly on the Itti model and its application in the image compression. The main work accomplished is as follows:First of all, Itti modeling process is discussed and based on which the research on focus transfer mechanism is improved. In dealing with images with Itti model, there may be such situation under which the same objects are paid attention many times (for example, the first time the upper half part of the object is noticed, while the second time the lower part is noticed). According to the theory that " There exist sensitive points in animal visions" , first, use local mean algorithm to figure out seven maximum points in the saliency map, and then, according to "adjacent point merge" method, make process on these seven maximum points, thus getting some attention focuses, and finally transfer the focuses of attention according to their significance. Results indicate that the improved model will only notice the same objects once in most cases. Therefore, this improved model is better than Itti model in identifying the integrity of objects.Secondly, a method to fix the interested region extraction based on the threshold value is put forward, in which a threshold value equivalent to 60% of focus significant value is set., and as long as the value of the pixels which are within a certain scope of attention focuses is greater than the threshold value, put these pixels to the interested region. The model has been experimented with some typical images; and the results are very satisfactory.Finally, the improved visual attention model is applied to image compression field. The corresponding image compression method can be obtained by compressing the background area with a high compression ratio instead of one or a few regions of interest distilled from the whole image , leaving them not being compressed or compressed with a low compression ratio. By combining JPEG2000 compression algorithm with this model, related experiments are held and the results are favorable.