Computational Visual Attention Model And Its Application In Image Classification

The mechanism of visual attention is a very important function for the human visual system, which enables us to quickly locate interesting regions or objects in complex scenes thereby perceiving external environment efficiently and accurately. Visual attention can be divided into two types based on different incentives:one is based on early visual features and bottom-up mechanism which is data-driven; the other is based on high-level vision features and top-down mechanism which is task-driven. In an image, there will at least one salient region and the other regions can be viewed as the background that human eyes are not interested and will generate interference in image description. The mechanism of visual attention will strengthen the salient region and at the same time weaken the background regions, so that we can just choose the useful information of the image for description and do not have to process the image without distinction.Due to the restrictions regulated by researches of physiology and psychology, existing visual attention models are mostly based on the bottom-up mechanism. And through the study on visual attention models, it finds that existing models may not be matched with the judgments of human eyes in describing images for some salient features may not be noticed in certain cases. For example, human eyes are likely to notice the regions with dramatic gray changes in an image but the existing visual attention models have not fulfilled that. According to the foresaid situation, the paper proposes a model which has combined visual attention and regional information entropy together, and the model has been applied into image classification.The algorithm used in this paper firstly determines the salient region of an image based on visual attention and regional information entropy, then uses SURF local feature detector to extract the image feature, and finally processes the extracted image features via the bag-of-word model, thus to realize an image description. This algorithm does not involve image semantics and only takes advantage of primitive features related with semantics, which can make up the deviations raised in image description with only bottom-layer features. The experimental results show that the proposed algorithm in the paper has achieved good results in multiple image libraries.