Approximately decoupled component supervision for salient object detection

Salient object detection (SOD) aims to find the most attractive object(s) in a scene. In recent years, SOD methods based on deep learning have become the mainstream. Existing methods mostly aggregate the multi-level features extracted by convolutional neural network (CNN) to model the object, or refine the boundary details of the object through multi-scale feature fusion. In this paper, the paradigm of approximately decoupled component supervision is proposed for SOD. Our insight is that the attractive performance of SOD requires explicit modeling of the body and edge of the object with different supervisions. Specifically, we first capture image features through foreground attention (FA) mechanism, cross-block semantic correlation aggregation (CBSC), and resolution-based feature integration (RFI) to make object parts more consistent and complete. Then the detailed edge is obtained by subtracting the body part from the complete mask. By explicitly sampling the body and edge pixels of the salient object, we further optimize the resulting body features and the residual edge features under the supervision of approximate decoupling. Benefiting from the abundant edge information and accurate location information, the framework with various backbone proposed by us can achieve better internal consistency and accurate boundaries of the object. Experimental results on six widely used benchmark datasets demonstrate the superiority and competitiveness of our approach in terms of four popular evaluation metrics. Moreover, the proposed method is an end-to-end saliency detection network without any pre-processing or post-processing.