SalNet: Edge Constraint Based End-to-End Model for Salient Object Detection

Abstract

Salient object detection is a fundamental task in computer vision and pattern recognition. And it has been investigated by many researchers in many fields for a long time. Numerous salient object detection models based on deep learning have been designed in recent years. However, the saliency maps extracted by most of the existing models are blurry or have irregular edges. To alleviate these problems, we propose a novel approach named SalNet to detect the salient objects accurately in this paper. The architecture of the SalNet is an U-Net which can combine the features of the shallow and deep layers. Moreover, a new objective function based on the image convolution is further proposed to refine the edges of saliency maps by using a constraint on the L1 distance between edge information of the ground-truth and the saliency maps. Finally, we evaluate our proposed SalNet on benchmark datasets and compare it with the state-of-the-art algorithms. Experimental results demonstrate that the SalNet is effective and outperforms several representative methods in salient object detection task.