Saliency Detection In Complex Scene And Its Application

With the rapid development of social media and e-commerce,the role of image in the exchange of information is growing.And the demand for image information processing is growing.It requires computers to be more intelligent and more autonomy to realize the image information processing.As a data storage format,the size of image data is too large.While the really useful information exists only in some part of the image.The human eyes can find objects or areas of interest in scenes quickly and spontaneously.This ability is able to assist the computers to find the effective representation objects or areas of the images efficiently.Saliency detection simulates the ability of the human eyes.Saliency detection draws some theoretical achievements in biology analysis of the human eyes and use computational methods to detect the objects or areas of interest from images.After a decade of development,saliency detection has gained tremendous development.A lot of saliency detection technologies show very high performance on some datasets.However,the current saliency detection technologies are still unable to cope with complex scenes and the case of multiple salient objects.This paper proposes two methods of these problems:(1)Saliency detection based on the object priori.First we use an object detection technology to extract the multi-class objects with a bounding box.Based on each objective prior,we conduct saliency detection.Then we evaluate all the saliency maps,and fuse the saliency maps.(2)Saliency detection based on eye fixation prediction.We first detect the regions which humans focus on,with the technology of eye fixation prediction.Then,we segment the foreground and conduct saliency detection based on the priori foreground.At the end,we study the saliency application on object recognition.We find that the current saliency application on object recognition confuse the information of foreground and background or ignore the information of background.We propose an approach of saliency based hierarchical fuzzy representation to integrate the information of foreground and background for object recognition,and we gain an improvement on performance.