Based On The Topological Properties Of The Visual Attention Model And Its Applications

Vision is one of the most important human being's perceptions. Attention selection is an important feature in visual perception. We can easily detect and recognize different objects in images and videos, but it is difficult for the traditional machine vision. Psychological research shows that, global topological perception of objects is perceived firstly in the process of human visual perception, followed by local features as brightness, color and motion. These local features are input parallel to visual neurons and processed synchronously. Attention selection plays a key role in the visual perception, which helps us focus and extract target regions with interest. Attention selection could be divided into two parts:bottom-up attention and top-down attention.We study the mechanisms of visual attention in this paper. Based on topological perception theory in psychology, topological properties are applied to bottom-up attention detection method. We also propose a new objective evaluation criterion for saliency maps, and apply saliency information in image segmentation. The main contribution of this paper includes:1. Topological properties are applied in an attention selection model. We propose a bottom-up attention detection method based on topological perception. In our model, the topological connectivity, color, brightness, motion and other features of visual information are extracted and input parallel to quaternion for processing. With hypercomplex Fourier transformation, the phase information of the original image could be obtained. After the inverse transformation to the spatial domain and filtering, the model's attention saliency map could be got finally. This method takes into account the important role of topological properties in visual perception and has a famous psychological theory as a basis. Our model can reflect attention saliency information distribution effectively.2. We introduce a new evaluation criterion for saliency map. The criterion is an objective evaluation without human beings participation. In bottom-up attention selection models, the results are displayed in saliency maps. It is always an important issue in this area that how to evaluate the quality of saliency maps effectively. Almost all of the existing evaluation methods require human beings participation. This kind of evaluation criteria has subjective factors inevitably without high reliability and persuasiveness. Our criterion is based on evaluation of the contribution of channels, as an objective evaluation, which does not need human beings participation. Based on the proposed evaluation criteria, we find some problems in the model introduced in the first part and improve. With adjusting the weight of the topological channel, the influence of that channel to the model could be reduced. The improved model can reflect the attention distribution more objectively and really.3. Attention saliency information is applied to image segmentation and achieves automatic segmentation of color images. Most of the existing color image segmentation methods rely heavily on manual marking, or parameter adjustment. We extract the image saliency information and improve the artificial mark of object and background regions to automatic mark by saliency maps. Combined with the mean shift and the maximal similarity based region merging segmentation methods, we introduce a new automatic color image segmentation method. The effect of our method is slightly improved, but the most important improvement is that it is an automatic color image segmentation method without needing human beings participation.