Saliency Detection with Stepwise Aggregation Network

Recently, benefiting from the progress of deep learning, salient object detection based on Convolutional Neural Networks (CNNs) has achieved significant improvement. For most methods that build upon Fully Convolutional Neural Network (FCN), how to utilize information from hierarchical features is a research hotspot. In this paper, we propose a novel aggregation module to integrate information through a step-by-step fusion mechanism. First, we construct a widely used encoder-decoder network as our framework and combine the features from encoder and decoder in the first step of our Stepwise Aggregation Module (SAM) to obtain a more comprehensive feature. Then, we feed the new feature map and the saliency map generated by previous stage into the Recurrent Integration Layer (RIL) to fetch information. Compared with previous methods, the proposed mechanism is simpler and more efficient. Experiments verify the effectiveness of our method. On five benchmarks, our model achieves the state-of-art performance.