Salient Object Detection Based on Unified Graph Neural Network Joint Learning

Abstract

In complex visual scene, the performance of existing deep convolutional neural network based methods of salient object detection still suffer from the loss of high-frequency visual information and global structure information of the object, which can be attributed to the weakness of convolutional neural network in capability of learning from the data in non-Euclidean space. To solve these problems, an end-to-end unified graph neural network joint learning framework is proposed, which realizes the joint learning process of salient edge features and salient region features. In this learning framework, we construct a multi-relations dynamic attention graph convolution operator, which captures non-Euclidean space global context structure information by enhancing message transfer between different graph nodes. Further, by introducing a graph attention fusion module, the full use of salient edge cues and salient region cues is achieved. Finally, by explicitly encoding the salient edge information to guide the feature learning of salient regions, salient regions in complex scenes can be located more accurately. The experiments on three public benchmark datasets show that our method has competitive detection results compared with the current mainstream deep convolutional neural network based salient object detection methods. More importantly, it uses fewer parameters and less computation, so it is a lightweight salient object detection model.