Abstract: Nowadays, deep convolutional neural networks (CNNs) have made significant improvement in detecting salient objects by integrating multi-level convolutional features or exploiting the advantage of dilated convolution. However, how to construct finer structure to produce effective features for saliency detection is still a challenging task. In this paper, we propose a novel deep learning based network by dynamically incorporating multi-level feature maps. The proposed Dynamically-passed Contextual Information Network (DCI-Net) can effectively control the information passage process, incorporate multi-scale context information and alleviate the distraction of background noise to improve the performance of saliency detection. Specifically, we first integrate an effective Passage Unit (PU) that progressively incorporates the low-level information with the high-level cues, in order to preserve the boundary details of salient objects. Second, a Spatial Pyramid Dilation (SPD) module is used to enhance the multi-scale feature representation by using multiple convolution dilations to handle the varied size of salient objects. Finally, we apply a Residual Attention module (RA) to further reinforce the saliency detection. Quantitative and qualitative experiments demonstrate the effectiveness of the proposed framework. Our method can significantly improve the performance based on five popular benchmark datasets.
0 Replies
Loading