Salient Object Detection via Light-Weight Multi-path Refinement Networks

摘要

Recently, deep learning-based saliency detection has achieved fantastic performance over conventional works. However, repeated subsampling operations in deep CNNs lead to difficulties for full resolution prediction and accurate prediction at boundaries of salient regions. In this paper, we use Lightweight Multi-Path Refinement Networks (RefineNet) for image saliency detection task, an encoder-decoder architecture that explicitly exploits all the information available along the down-sampling process to enable full resolution prediction using long-range residual connections. The squeeze and excitation residual network (SE-ResNet) is adopted as our baseline network to better extract multi-level and multi-scale feature maps according to the resolutions. Furthermore, we proposed our end-to-end network architecture and reduced the parameters by more than 50% while maintaining the same performance. We carry out comprehensive experiments and set new state-of-the-art results on four public datasets. By contrast, our method is highly competitive.