Computational Methods Of Visual Saliency Based On Scene Structure

Visual attention,as an important research topic of biological vision and computer vision,has attracted many experts and scholars in the fields of psychology,neuroscience and computer vision since it was proposed.It has been found that when the human brain,and visual system process visual scene images,they don't treat all information equally but tend to distribute more attention to certain regions or objects(such as the order and duration of observation),which we call salient objects or regions.The algorithms of salient objects and salient regions detection are widely used in engineering,such as intelligent cameras,intelligent traffic systems,etc.With more scholars and experts in neurophysiology and anatomy devoting to the research on the visual information processing mechanism,the achievements of the biological visual mechanism have been gradually increasing,and the understanding of the biological visual information processing mechanism has been becoming broader and deeper.This inspires us to develop image processing algorithms guided by biological visual mechanism,so as to produce a more intelligent visual processing system.At present,the research on visual attention mainly includes the prediction of eye fixation points and the detection of significant objects.This thesis mainly researches the human eye fixations of the visual attention model.Based on the visual perception mechanism and the gestalt cognitive psychology,we explored the visual saliency related to the low-level,mid-level,and high-level visual features to building the algorithm for saliency detection to solve the related problems from the following two aspects.(1)In this thesis,a method of saliency detection based on scene contour is proposed,which can predict the visual saliency of natural images by combining the low-level and middle-level visual clues of scene images.First,three mid-level cues based on gestalt principles are defined,including contour density,closure,and symmetry to describe potentially saliency regions.In addition,we use the classic bottom-up approach to generate the underlying saliency map.Finally,the low-level feature of the natural image is combined with the mid-level feature of the corresponding contour to generate a saliency map based on the scene contours.The experimental results show that the middle-level feature based on contour can significantly improve the performance of visual saliency prediction of the bottom-up model.(2)Different objects have different probabilities in each region of the scene.Therefore,a method for detecting the saliency of specific objects(oil paintings)based on indoor scenes is proposed in order to further explore the relationship between visual saliency and scene structure.First of all,the deep learning method is used to generate a coarse-grained semantic segmentation structure of the interior image for encoding the indoor scene.Then,the probabilities of a specific object appearing in different regions of the scene in the database are calculated and used as prior information.Finally,Bayes' theorem is utilized to calculate the conditional saliency of specific objects which is to generate the final visual saliency map.Experimental results show that scene structure plays an important role in visual search and visual saliency detection.