Fixed Feature With Attention Based on Recurrent Criss-Cross Salient Object Detection Network

Salient object detection (SOD) is a basic research direction in the computer vision task. Its main work is to simulate the human visual system by establishing a visual-aware model. However, most of existing models have suffered from two challenges. First, edge-aware models only use edge feature to improve segmentation features but lack of utilizing the complementarity of them. Second, the problem of losing part of information in the process of extracting features in Convolutional Neural Networks (CNN) leads to the incompleteness of final result. In this paper, we proposed a criss-cross refined salient object detection method based on hybrid attention to solve these problems. Specifically, we applied a Criss-Cross Attention module (CCA) to capture useful contextual information of surrounding pixels from long-range dependencies effectively. Fully considering the complementary between salient detection and edge information, our model simultaneously refines multilevel features of SOD and edge detection by overlapping a Cross Refinement Unit (CRU). And finally we used a U-Net to output the refined result. Extensive experiments conducted on five benchmark datasets demonstrate that our method can significantly improve the performance of low-contrast scenes, optimize the problem of partial missing and error detection of salient objects. Besides, our method achieved state-of-the-art performance in three different indicators.