边缘信息增强的显著性目标检测网络

Edge Enhancing Network for Salient Object Detection

摘要:

针对显著性目标检测任务中识别结果边缘模糊的问题，提出了一种能够充分利用边缘信息增强边缘像素置信度的新模型。该网络主要有两个创新点：设计三重注意力模块，利用预测图的特点直接生成前景、背景和边缘注意力，并且生成注意力权重的过程不增加任何参数；设计边缘预测模块，在分辨率较高的网络浅层进行有监督的边缘预测，并与网络深层的显著图预测融合，细化了边缘。在6种常用公开数据集上用定性和定量的方法评估了该模型，并且与其他模型进行充分对比，证明设计的新模型能够取得最优的效果。此外，该模型参数量为30.28 M，可以在GTX 1080 Ti显卡上达到31 帧·s^-1的预测速度。

Abstract:

Aiming at the problem of blurred edges in salient object detection， this paper proposes a new method that can fully utilize edge information to enhance the confidence of edge pixels. First， the triple attention module is introduced， which uses the characteristics of the predicted saliency map to directly generate foreground， background and edge attention， and the process of generating attention weights does not add any parameters. Next， the edge prediction module is introduced， which performs supervised edge prediction in the shallowest layer of the network with the biggest feature map， and fuses the predicted edge with the saliency map to refine the edges. Finally， the model is qualitatively and is quantitatively evaluated on six commonly used public datasets， and fully compared with other models， which proves that the proposed model can achieve the best results. The method proposed in this paper has 30.28 M parameters， and can predict saliency maps at 31 frames per second on GTX 1080 Ti graphics card.