Recently, the convolutional neural network (CNN) has achieved great progress in many computer vision tasks including object detection, image restoration, and scene understanding. In this paper, we propose a novel CNN-based saliency detection method through dense recurrent connections and residual-based hierarchical feature integration. Inspired by the recent neurobiological finding that abundant recurrent connections exist in the human visual system, we firstly propose a novel dense recurrent CNN module (D-RCNN) to learn informative saliency cues by incorporating dense recurrent connections into sub-layers of convolutional stages. Then we present a residual-based architecture with short connections for deep supervision which hierarchically combines both coarse-level and fine-level feature representations. Our end-to-end method takes raw RGB images as input and directly outputs saliency maps without relying on any time-consuming pre/post-processing techniques. Extensive qualitative and quantitative evaluation results on four widely tested benchmark datasets demonstrate that our method can achieve more accurate saliency detection results solutions with significantly fewer model parameters.
- Convolutional neural network
- Deep supervision
- Hierarchical feature fusion
- Recurrent convolutional layer
- Salient object detection