Visual Saliency Detection Based On Eye Tracking Prior Knowledge

Studies in biology show that there is a selective visual attention mechanism in human visual system,when seeing a scene at first glance,human can quickly focus on the region with largest information.The visual saliency detection aims to simulate this mechanism of human visual system to make the computer more rapid and intelligent when processing a scene picture or video.How to extract the most attractive area from an image automatically and accurately is the most significant goal of saliency detection.As an independent research direction,saliency detection provides great help for analyzing and understanding the image content.Taking saliency detection for the image in the preprocessing stage firstly can greatly improve the efficiency of subsequent processing.The saliency detection is also very meaningful for some other areas.The work firstly reviews the current development of saliency detection and introduces some classic model,analyzes the pros and cons of saliency detection algorithm which based on learning.After that,the paper proposes a saliency detection algorithm based on sparse constraint of image features from eye tracking data.The algorithm firstly establish a feature pool which include a variety of image features,then suppose that image's saliency map can be represented by a linear combination of features in the feature pool and learn the weight parameters of this linear combination from the prior knowledge of eye tracking database using linear regression with sparse constraint.Compared with other saliency detection algorithm based on learning,this algorithm mining the information of eye tracking data more fully,it not only generates the integration parameters automatically,but also achieve the purpose of feature selection and redundant information removal.The experiments shows that the saliency detection model constructed by this algorithm gets a very good performance in detection accuracy and detection efficiency.The work also proposes a visual feature extraction algorithm based on eye tracking data.Using the image patch as basic operation unit,the algorithm firstly construct the image patch feature by using color,intensity and orientation.Than use the K-means algorithm to construct an image word mapping collection and calculate the level of saliency.At last,for the image which is waiting detection,after the segmentation and the classification of image patch,the algorithm calculate the saliency prior feature according to the mapping collection.The experiments shows that when adding the new feature to the saliency detection model,the detection result improved significantly.