Saliency Detection Based on Deep Learning and Graph Cut

Abstract

In this paper, we propose a new saliency detection method based on deep learning detection models and graph-cut partitioning methods. First, we present a full convolution deep neural network saliency detection model. By training a binary classification, full convolution, deep neural network, the image is divided into foreground and background. Given the initial image, the output is the coarse-grained saliency map with foreground and background. Second, we present a saliency detection method based on graph community partitioning. Using the superpixels algorithm, the initial image is transformed from pixel level to region level, and the super-pixel region similarity matrix is constructed. The similarity matrix is automatically divided into areas by the community-partitioning algorithm, and the segmentation map is obtained. The saliency values of each region are computed by the Gauss kernel function. Lastly, we combine the high-level global saliency map of deep learning with the low-level local saliency map of graph partitioning cut, and realize the saliency detection task of the original image. To illustrate the effectiveness of the proposed method, we also compare the method with the seven popular saliency detection algorithms and evaluate them on four public image databases. A large number of experimental results demonstrate that this algorithm has obvious superiority in comparison with existing algorithms.