Graph Saliency Network: Using Graph Convolution Network on Saliency Detection

Abstract

Saliency detection is to detect the unique region of an image that may attract human attention. It is widely used in image/video segmentation, image enhancement, and image compression. Conventionally, saliency detection problem was solved by graph-based method cooperate with low-level features and heuristic rules. Recently, the convolutional neural networks (CNNs) based methods have been thrived in computer vision area and graph convolutional networks (GCNs), which are extended from the CNN, have been used in many graph data representations and also shown promising result in node classification problem. We proposed a novel saliency detection neural network model called the Graph Saliency Network (GSN), which use the Graph Convolutional Network as main architecture and the Jumping Knowledge Network as our backbone. For the graph creation, the Region Adjacency Graph is adopted as the image-graph transformation in the proposed architecture to propagate information through edges from the spatial boundary. We also revisit several graph-based saliency detection methods for our node feature representation. The propagation model of the GSN maintain the spatial relation of the CNN with a more flexible way and has less parameters to be optimized than the CNN from the advantage of information compression in superpixel and graph. Simulations showed that, using the proposed GCN- based model together with low-level features and heuristic rules, a saliency detection result with very less mean absolute error (MAE) can be achieved.