A novel dynamic graph evolution network for salient object detection

The advanced deep convolution neural networks (CNNs) based salient object detection (SOD) models still suffer from the coarse object edge. The traditional graph-based SOD models can preserve the object edge well benefitting from the superpixels, but they perform weaker in highlighting the whole object compared to recent deep learning models. To tackle this problem, we attempt to find a new way to address this issue under the framework of graph convolution networks (GCNs). Specifically, we first model the image as a set of superpixels and construct the graph structure by connecting the k nearest neighbors for each node. For the connected nodes, rather than only leveraging on the predefined edges, we propose a multi-relations edge convolution operation, expecting to learn multiple implicit relations in the pair-wise nodes and aggregate the information from their neighbors relying on the learned edges. Then, a channel-wise attention operation is also proposed to boost the intra-node massage propagation across channels within the same node. In order to optimize the graph structure from layer to layer, we learn a new metric to re-measure the distance between any pair of nodes, and the graph structure evolves dynamically by recomputing the k nearest neighbors in the saliency feature space at different layers. Finally, a residual structure is applied to enable our graph network to go as deep as CNNs models. The graph nodes (superpixels) inherently belonging to the same class will be ideally clustered together in the learned embedding space. Experiments show that this work is a good practice for designing GCNs for image SOD and achieves the comparable performance with the recent state-of-the-art deep CNNs-based models.