PROJECT TITLE :

Reverse Attention-Based Residual Network for Salient Object Detection

ABSTRACT:

Recent advances in salient object detection have been made thanks to the rapid development of deep convolutional neural networks, particularly fully convolutional neural networks (FCNs). It is still difficult to produce high-resolution saliency maps using these FCNs based approaches because of the hefty model weights. For salient object detection, we offer a compact, efficient, and high-accuracy deep network in this research. A multi-scale context module and hand-crafted saliency priors are the first two methods we suggest for making first predictions. To further reduce the number of convolutional parameters required, we solely use residual learning to learn the residual in each side-output, resulting in a highly compact and highly efficient model. A new top-down reverse attention block is then designed to steer the above side-output residual learning. Its side-output feature is erased using the current projected salient regions, and the missing object components and details are efficiently learned from these unerased regions, resulting in more complete detection and high accuracy. Seven benchmark datasets reveal that the proposed network outperforms state-of-the-art techniques in terms of simplicity, compactness, and efficiency, as demonstrated by extensive experiments.