Research Of Image Saliency Prediction Method In Social Scene

Human visual attention,the ability to quickly and accurately catch small parts of the scene for further visual related processing,plays an important role in visual perception.Transferring such a selective attention mechanism to image processing field is a straightforward approach and gives us a new insight for interpreting the saliency in a vision.According to this view,saliency prediction has been served as a preprocessing step in almost image related works and obtained some satisfactory results.There are many saliency models have been proposed with bottom-up features and gained satisfactory fixation prediction of free-viewing.However,due to the different visual task or complex scene,the computational cognitive model is far from the observer's eye saccade behavior.This article draws on the theoretical achievements of the visual attention mechanism in cognitive psychology and neurobiology,then combines the eye movement experiment data analysis conclusions to study the image saliency prediction method in social scenes.The main works are as follows:(1)We construct a dataset named Gaze-in-Scene-232(GS232),which consists of 232 social images,and for each image,eye tracking data is obtained from 30 subjects with three different visual tasks.According to eye tracking data,we find that face and gaze can quickly attract the observer's interest,especially for gaze-following visual task.(2)We extract top-down features with proposed Pyramid-Box and Gaze Net.And we propose an optimization algorithm for saliency prediction with GMM(Gaussian mixture model),which integrates both bottom-up and top-down features.(3)We explore a strategy for generating scanpath under different tasks for social scene image datasets,and propose a scanpath generation algorithm based on the prior information of social scene.