Image Saliency Detection Algorithm Based On Foreground Prior And Saliency Enhancement

The research on computer vision has developed vastly.Saliency detection is one of the important methods which is used to preprocess images and videos.The main purpose of saliency detection is to find the proper features to highlight the foreground area from background area correctly.However,we still don't know how the human visual system works on image saliency detection.Scientists can only simulate the visual system by using prior knowledge and hypothesis model.However,as the foreground and background of images can be very complicated,many saliency detection algorithms cannot achieve ideal results.Saliency detection methods can be divided into two classes.By utilizing contrast prior and boundary prior,bottom-up algorithms can construct mathematical model to compute saliency value of every pixels in an image.Top-down methods try to learn the right way to extract features and construct reasonable models with plenty of labeled data.We propose two novel algorithms on these two aspects.On the bottom-up aspect,we propose a novel method based on the foreground area ratio prior.We first detect the corners in one image with corner detection method.Then we build a convex hull with these corners which will encompass most of the foreground area.With this convex hull information,we compute the saliency value of each superpixel by the foreground area ratio of each superpixel in convex hull.Because this method can roughly decide the position of foreground area,it will suppress the background more effectively compared to boundary prior.Moreover,we propose a iterative optimization method based on graph regularization.This optimization method promotes the saliency value of salient object,meanwhile suppressing the background saliency value.On the top-down aspect,we propose a combined saliency enhancement method based on fully convolutional network.Many methods have proved that we can not make a further progress of saliency detection only with priors.Hence,researchers try to promote the saliency value with machine learning method.In our work,we propose a saliency enhancement method with deep convolutional network.We have proved that our methods have shown good performance which is on the same level with other state-of-art methods on four standard datasets.