Feature Recalibration Network for Salient Object Detection

Abstract

Learning-based models have demonstrated the superiority of extracting and aggregating saliency features. However, we observe that most off-the-shelf methods mainly focus on the calibration of decoder features while ignore the recalibration of vital encoder features. Moreover, the fusion between encoder features and decoder features, and the transfer between boundary features and saliency features deserve further study. To address the above issues, we propose a feature recalibration network (FRCNet) which consists of a consistency recalibration module (CRC) and a multi-source feature recalibration module (MSFRC). Specifically, intersection and union mechanisms in CRC are embedded after the decoder unit to recalibrate the consistency of encoder and decoder features. By the aid of the special designed mechanisms, CRC can suppress the useless external superfluous information and enhance the useful internal saliency information. MSFRC is designed to aggregate multi-source features and reduce parameter imbalance between saliency features and boundary features. Compared with previous methods, more layers are applied to generate boundary features, which sufficiently leverage the complementary features between edges and saliency. Besides, it is difficult to predict the pixels around the boundary because of the unbalanced distribution of edges. Consequently, we propose an edge recalibration loss (ERC) to further recalibrate the equivocal boundary features by paying more attention to salient edges. In addition, we also explore a compact network (cFRCNet) that improves the performance without extra parameters. Experimental results on five widely used datasets show that the FRCNet achieves consistently superior performances under various evaluation metrics. Furthermore, FRCNet runs at the speed of around 30 fps on a single GPU.