Abstract: Salient object detection is a popular research area in computer vision. Numerous deep learning-based methods have been proposed, and have shown great potential. However, there are some problems with the detection, such as false detection or fuzzy edges. To address these problems, this paper proposes a novel dual-stream fusion and edge-aware network for salient object detection. By utilizing the multi-scale information and features aggregating, the model can obtain the fine-grained results. Firstly, the input image is modified into two different scales to feed to the encoder of the network separately. In this way, the model can extract the abundant multi-level features from the dual-feature stream. Secondly, in the stage of the decoder, the dual-stream features are gradually integrated from top to bottom, to generate the coarse-to-fine salient map. Moreover, a dual-path edge-aware structure is designed, which can generate the subtle boundary by fusing the semantic contextual information. Finally, the edge refinement module is used to improve the salient object boundaries, and obtain the final results. Extensive experiments are conducted on five public datasets, the results show that the proposed method achieves higher accuracy in terms of structural similarity (Sm), and generates high quality salient maps with more complete objects and distinct edges.

Key words: salient object detection, full convolutional neural network, multi-scale learning, dual-stream fusion, edge-aware

摘要： 显著性目标检测是计算机视觉领域的热门研究方向之一，许多基于深度学习的检测算法虽然已经取得了显著的成果，但是仍然存在待测目标漏检误检和边界模糊等问题。针对这些问题提出了一种基于双特征流融合和边界感知的目标检测算法，通过改变输入图像尺寸来丰富多尺度信息，并自顶向下逐层聚合特征得到精细的预测结果。首先将输入图像调整为两种不同分辨率分别送入编码器，提取丰富的多层级特征形成双特征流；其次将双特征流自顶向下逐层融合，生成由粗到细的显著图；最后构建了边界感知结构，凭借上下文语义信息的指导生成精细的物体轮廓。在五个公开数据集上进行了大量实验，实验结果表明，所提算法在结构相似性[(Sm)]等多个指标上取得了更高的检测精度，生成的显著图目标完整且边缘清晰。

关键词: 显著性目标检测, 全卷积神经网络, 多尺度学习, 双特征流融合, 边界感知