Saliency Detection With Region Regression

Saliency detection refers to extract the significant areas in an image by simulating human visual features.It is an important pre-processing method for a computer to further understand and process an image.Although there are many researches on saliency detection,there is no clear explanation on how to capture saliency region in human vision system.Because a large number of images also contain rich foreground area and background area information,saliency detection still has great room for progress.Especially when the image contains a complex scene,how to extract salient objects efficiently and accurately is still a big problem to be solved.With the development of convolutional neural network,saliency detection has made great progress.Computer no longer need to provide a lot of prior information,because neural network can extract the significance feature from batch training data.However,many saliency detection methods based on deep learning have been able to find the salient region of the image accurately,but they sometimes mistakenly recognize non-salient objects as salient objects,especially the image with a very complex background.In order to solve this problem,this paper propose a two-stage image saliency detection method.In the first stage,the algorithm should separate the salient region from the original image as much as possible.In this paper,a neural network is utilized to regress the minimum saliency region(RMSR)which contains all salient objects and remove the background as much as possible.Through it,the problem of salient object location in complex scenes can be solved.Then in the second stage,in order to fuse multi-level features as much as possible,spiral shared neural network is proposed to complete the pixel level saliency detection on the result of RMSR.Through a two-stage algorithm model,computer can accurately extract salient objects from an extremely complex original image.Experimental results on four public datasets show that this model is effective over the state-of-the-art approaches.