Design Andimplementation Of A Biological Plausibilityvisualattention Model Infrequency Domain

The selective mechanism of human visual attention can make sure that we can rapidly and precisely locate the most important region when face complex scene and deal with it preferentially and neglect redundant information. The data which most of image processing tasks need is just a part of the whole image, but existing methods handle all image data in a non-hierarchical way and waste a lot of computing time and space. Therefore, it is necessary to introduce the selective mechanism of human visual attention into image processing, besides, it can be widely used in object recognition and detection, image classification and retrieval, and so on. So it is of great significance in researching on it.Visual attention models in frequency domain have been a topic in the past years. However, they are all lack of biological theory and cannot been accepted, its biological plausibility needs to be improved and enhanced. Based on the analysis of existing visual attention models in frequency domain, we design a biological plausibility model in frequency domain, and apply it into image classification. The main works of our paper are as follows:(1) The visual attention mechanism and the status of visual attention models are reviewed in detail. Besides, we give a brief introduction of the basic theory of quaternion and analyze the advantages and disadvantages of current visual attention models in frequency domain.(2) Aiming at addressing the biological plausibility of models in frequency domain, the HFT(Hypercomplex Fourier Transform) model was improved in three aspects in this paper in order to make it receivable. Firstly, we choose CIE Lab color space which coincides with human visual perception instead of RGB color space. Secondly, the imaginary coefficients of a quaternion image imitate the center excitedsurround inhibited characteristic of human visual neurons. At last, the final saliency map was obtained by integrating the saliency maps of different scales according to a certain rule. The improved HFT model was tested in public image databases, and the experimental results show that the saliency map achieved by the improve HFT is closer to human visual system and the areas under the receiver operating characteristic curve of the improved HFT model in Bruce and Judd databases increase 0.97% and 6.33%, respectively..(3) A method of image classification was proposed based on the visual saliency. Firstly, we use the improved HFT model to achieve the salient region; then the features of the salient region were extracted in order to accomplish image classification. In the process of feature extraction, we introduce the image time icon computed by pulse coupled neural network. The experimental results demonstrate that the image classification precision rate of the proposed method based on salient region increases 4.3% compared with the method based on the whole image.The work we study begins with the biological plausibility of visual attention models in frequency domain, so it is of innovativeness and challenging. The improved HFT model can meet the requirement of biological plausibility and be applied into image classification as a new idea. The proposed model has certain theoretical significance and practical value.