Image Saliency Detection with Low-Level Features Enhancement

Abstract

Image saliency detection has achieved great improvements in last several years as the development of convolutional neural networks (CNN). But it is still difficult and challenging to get clear boundaries of salient objects. The main reason is that current CNN based saliency detection approaches cannot learn the structural information of salient objects well. Thus, to address this problem, this paper proposes a deep convolutional network with low-level feature enhanced for image saliency detection. Several shallow sub-networks are adopted to capture various low-level information with heuristic guidance separately, and the guided features are fused and fed into the following network for final inference. This strategy can help to enhance the spatial information in low-level features and further improve the accuracy in boundary localization. Extensive evaluations on five benchmark datasets demonstrate that the proposed method outperforms the state-of-the-art approaches in both accuracy and efficiency.