PROJECT TITLE :

Residual Learning for Salient Object Detection

ABSTRACT:

Most recently developed salient object detection deep learning algorithms use multi-scale methodologies and fully convolutional neural networks to enhance performance (FCNs). Predictions at different scales are combined to get a final conclusion. However, there are significant issues with the current multi-scale methods: 1) it is difficult to learn discriminative features and filters directly to regress high-resolution saliency masks for each scale; 2) rescaling the multi-scale features could bring in many redundant and erroneous values, weakening the network's representational capabilities. To gradually improve the coarse prediction, we present a residual learning technique in this study. On a concrete level, we learn to forecast residuals rather than the fine-resolution result directly at each scale in order to correct for differences between the coarse saliency map and the scale-matching ground truth masks. The coarse prediction is generated using a Dilated Convolutional Pyramid Pooling (DCPP) module, and the residual learning is guided by several new Attentional Residual Modules (ARMs). "R 2 Net" stands for Residual Refinement Network. On five publicly available benchmark datasets, we compare the proposed method's performance to that of other recent state-of-the-art methods. No additional post-processing is required for our R 2 Net, which runs at a real-time pace of 33 frames per second on a single GPU.