| 引用本文: | 张德华,李俊豪,张静凯,肖启阳.高效通道注意力和特征融合的协同显著性检测算法[J].哈尔滨工业大学学报,2022,54(11):103.DOI:10.11918/202109111 |
| ZHANG Dehua,LI Junhao,ZHANG Jingkai,XIAO Qiyang.Co-saliency detection algorithm with efficient channel attention and feature fusion[J].Journal of Harbin Institute of Technology,2022,54(11):103.DOI:10.11918/202109111 |
|
| 摘要: |
| 针对现有的协同显著性检测算法在多显著目标复杂场景下表现不佳的问题,提出了一种基于高效通道注意力和特征融合的协同显著性检测算法。首先,检测算法利用预训练的深度卷积神经网络对场景进行多尺度特征的提取,结合边缘显著信息设计了显著性语义特征提取模块,以避免全卷积神经网络导致边缘信息的缺失；其次,通过内积基本原理得到组内图片间的关联性信息并根据其关联程度进行自适应加权,结合高效通道注意力层设计了协同特征提取算法；最后,为了将各级高层语义特征经过协同显著性特征提取之后的结果与浅层次的特征进行融合,并实现对预测结果进行多分支同步监督,设计了基于高效通道注意力的特征融合模块。通过对3个经典的数据集进行测试,并与6种现有的协同显著检测算法进行对比,结果表明本文所提算法提高了复杂场景中图像的协同显著性检测的精度以及边缘信息的丰富程度,并具有更优的协同显著性信息检测性能；通过消融实验进一步验证了所提设计算法各个模块的有效性和必要性。 |
| 关键词: 深度卷积神经网络 协同显著性检测 多尺度特征 特征提取 特征融合 注意力机制 |
| DOI：10.11918/202109111 |
| 分类号:TP212 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(6,1); 河南省科技厅科技攻关项目(222102220028); 河南省高等学校重点研究计划(20A5,2A416004); 河南大学一流学科培育项目(2018YLTD04); 河南省青年人才托举计划(2021HYTP014) |
|
| Co-saliency detection algorithm with efficient channel attention and feature fusion |
|
ZHANG Dehua1,LI Junhao1,ZHANG Jingkai2,XIAO Qiyang1
|
|
(1.School of Artificial Intelligence, Henan University, Zhengzhou 450046, China; 2.Miami College, Henan University, Kaifeng 475004, Henan, China)
|
| Abstract: |
| Considering the poor performance of existing co-saliency detection algorithms in multiple salient object complex scenarios, a co-saliency detection algorithm with efficient channel attention and feature fusion was proposed. Firstly, the pre-trained deep convolutional neural network was adopted to extract multi-scale features of the images, and a saliency semantic feature extraction module with edge saliency feature was designed to avoid the lack of edge information caused by fully convolutional neural networks. Secondly, the association information between images in the group was obtained based on the inner product principle, and adaptive weighting was carried out according to the association degree; a collaborative feature extraction algorithm was designed in combination with the attention layer of efficient channel. Finally, a feature fusion module based on efficient attention layer was designed, so as to fuse the results of co-saliency feature extraction at high-level semantic features with low-level features, and supervise the predictions with multi-branches simultaneously. Three classic datasets were tested, and six existing collaborative saliency detection algorithms were compared with the proposed algorithm. Results show that the proposed algorithm not only improved the accuracy of collaborative saliency detection and the richness of edge information in complex scenarios, but also had better performance of collaborative saliency detection. The effectiveness and necessity of each designed module were further verified by ablation experiments. |
| Key words: deep convolutional neural network co-saliency detection multi-scale feature feature extraction feature fusion attention mechanism |
