Saliency Detection With Recurrent Fully Convolutional Networks

Salient object detection is to highlight the most outstanding or the most important region of an image for further usage.It is well known that the visual system of humans and other primates is similar.The important object regions have some properties without other tasks or interference,for example,a flower in a scene surrounded by grass is salient,so is a man standing in the desert.These objects has some common properties,such as higher contrast degree with background,smaller region in the scene and with closed graphical boundaries.For the simple scenes above,there are many efficient algorithms to solve them.However,for more complicated situation,none of them can achieve desired result.Most of these models is sensitive to noise of background.Whereas human visual system can focus on the salient objects rapidly and accurately.Based on this point of view,more robust model is necessary to adjust to intricate scenes.This paper proposes a novelty saliency detection model based on recurrent fully convolutional networks.Firstly,traditional contrast features are utilized to produce an initial map,including color and intensity.Next,the image and initial map are sent to the deep model for training.An image is input into the network,encoding with convolutional and pooling layers and decoding with deconvolutional layers.The network outputs an saliency map with the same size as the image.Meanwhile,the predicted map is sent to the input of the network with the recurrent structure to replace the saliency map of last time.To solve the problem of object size,the multi-scale structure is adopted to fuse the results of layers inside the network.The usual training method can fall into local minimum for this model,so the two-stage training strategy is proposed,including pre-training and fine-tuning.This strategy can make full use of the supervision information of segmentation data set,so can mitigate this problem properly.The proposed model is evaluated and compared with eleven state-of-the-art methods on four publicly standard salient object detection databases.Experimental results show that the proposed method comfortably outperforms other state-of-the-art methods in terms of PR curve,F-measure and visual quality.