Boosting Feature-Aware Network for Salient Object Detection

Jianwei Zheng, Yubin Gu, Yuchao Feng, Jinshan Xu, Meiyu Zhang

Abstract: Deep convolutional neural networks have demonstrated competitive performance in salient object detection. To capture more precise saliency maps, recent approaches mainly concentrate on aggregating deep features from convolutional networks and introducing edge supervision for a guarantee of compact targets. Though significant progress has been accomplished, the problems of low-contrast by targets against backgrounds and the inconsistency of object size are still challenging to tackle. To relieve these troubles, we propose a boosting feature-aware network (BFANet) following a dual-stream architecture, including an object sub-network for salient objects detection and a boundary guidance sub-network supervised by enhanced labels for edge detection. Specifically, for the interesting yet low-confidence areas, we mold a boosting feature-aware module (BFAM) hammering at suppressing the background while highlighting the weak features. In addition, considering the different responses of channels to output, we present a weighted aggregation block (WAB) to strengthen the significant channel features and recalibrate channel-wise feature responses. Extensive experiments on five benchmark datasets demonstrate that our proposed model outperforms most other state-of-the-art methods.