Semantical Knowledge Guided Salient Object Detection with Multiple Proposals

Abstract

In recent years, the area of salient object detection has developed rapidly duo to the revival of deep learning techniques, especially the emergence of deep Convolutional Neural Networks (CNNs), which has greatly boosted the detection result. Although CNNs can be used to perceive the salient objects, it is difficult to work for images with complex background, which is also a major problem faced by most work. In this paper, we introduce a salient object detection method using high-level features with semantic meaning. In order to obtain the accurate region and edge of all the salient objects in an image, our model has two designs: (1) utilizing multiple proposals as the semantical knowledge prior to enhance the power of locate objects, that are most likely to cover the entire salient regions of an image, and (2) using several attention modules to improve the representation ability of our model, and using abundant low-level feature information extracted by the encoder of the network to assist its decoder to obtain the precise saliency map relatively. In addition, our model can make full use of multi-level features and semantical knowledge, so the saliency map we got is very excellent. The experiments shows that our approach achieves state-of-the-art performance on four public benchmarks, and produces significant improvements over existing well-known methods.