Study On Predicting Human Gaze With Information From Multi-channel

There is a visual attention mechanism in the human visual system to give priority to dealing with some parts of the image,and then dealing with or ignoring other parts.This visual attention mechanism solves the contradiction between the amount of input information and processing information.Saliency detection is designed to simulate the human's visual attention mechanism and select certain regions in the complex images which can attract visual attention.Saliency detection technology has a huge development potential and broad application prospect in the practical engineering application,such as image segmentation,adaptive image compression,content based image editing,object detection and recognition,and image retrieval.Through a large amount of saliency research,we analyzed a large number of algorithms,and summarized the existing eye fixation detection techniques.Combined with the current technology,we proposed a new algorithm in the paper.Considering the diversity of natural scene images and the attention mechanism of human eyes,we adopted a multi-channel detection framework.Different channels corresponded to different levels of information.In the low-level information channel,we used the color information to detect regions with strong contrast to other regions in the image.In the salient object detection channel,the multiple instance learning(MIL)algorithm was combined with supervised learning classification algorithm to detect objects which can attract human's attention.The multiple instance learning algorithm can avoid the fuzziness of sample selection in supervised learning,and there was a complementary relationship between them.In addition,depending on the content of the image,the observer is more or less inclined to look at the center of the image.With this tendency in mind,a center biased channel was added.In order to fuse these information more scientifically,a deep neural network was trained to determine the reliability of the detection results for each channel.The weight of each channel detection result was determined by the output of the network.Finally,the proposed algorithm added three semantic detectors to improve the detection effects further.The main contribution of this paper:(1)we adopt a multi-channel detection framework,utilize the high,medium,low levels of information to detect the eye fixation.We archive good detection results in the detection of images which contain a variety of contents;(2)multi instance learning(MIL)algorithm is applied to the eye fixation detection for the first time,to determine whether the objects in the image is salient,solving the problem of fuzzy sample selection,which makes up the problem of inaccurate sample selection in supervised learning.(3)a new adaptive weight fusion method based on deep learning framework is proposed to integrate information at different levels.The algorithm was evaluated in three international databases including MIT1003,MIT300,and Toronto120 and compared with the six excellent algorithms in the world.Experiments show that our algorithm is beyond other algorithms in each evaluation standard,and leads good detection effect.