Salient object detection with non-local feature enhancement and edge reconstruction

The salient object detection task based on deep learning has made significant advances. However, the existing methods struggle to capture long-range dependencies and edge information in complex images, which hinders precise prediction of salient objects. To this end, we propose a salient object detection method with non-local feature enhancement and edge reconstruction. Firstly, we adopt self-attention mechanisms to capture long-range dependencies. The non-local feature enhancement module uses non-local operation and graph convolution to model and reason the region-wise relations, which enables to capture high-order semantic information. Secondly, we design an edge reconstruction module to capture essential edge information. It aggregates various image details from different branches to better capture and enhance edge information, thereby generating saliency maps with more exact edges. Extensive experiments on six widely used benchmarks show that the proposed method achieves competitive results, with an average of Structure-Measure and Enhanced-alignment Measure values of 0.890 and 0.931, respectively.