Deep Saliency Detection via Spatial-Wise Dilated Convolutional Attention

Abstract Saliency detection aims to highlight the area which significantly attracts human attention and stands out in an image. In recent years, deep learning-based saliency detection has achieved fantastic performance over conventional works while still facing huge challenges in multi-features fusion and the enlargement of the receptive field. Current top-performing saliency detectors on the basis of FCNs benefit from their powerful feature representations but suffer from high computational costs due to the integration of multi-scale features without distinction. So in this paper, we propose a novel and simple network, the DCAM, based on attention mechanism with dilated convolutions(DAM), incorporating multi-scale features with enlarged receptive field. Specifically, we apply DAM to guide each side output respectively which selectively emphasizes the significant regions, thus efficiently enhancing the representation ability of each layer. Our spatial attention module helps us looking for areas in the image that have a greater impact and give them a higher weight.Besides, we adopt FPN to integrate features adjacent to each other layer and a CRF scheme for refining saliency results. Experiments on five benchmark datasets demonstrate that the proposed approach performs favorably against five state-of-the-art methods with a fast speed (56 FPS on a single GPU).