基于双注意力模块的FDA-DeepLab语义分割网络  

FDA-DeepLab semantic segmentation network based on dual attention module

作　　者：张小国[1] 丁立早 刘亚飞 郑子豪 王庆[1] Zhang Xiaoguo;Ding Lizao;Liu Yafei;Zheng Zihao;Wang Qing(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China;College of Software Engineering,Southeast University,Suzhou 215123,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096 [2]东南大学软件学院,苏州215123

出　　处：《东南大学学报（自然科学版）》2022年第6期1145-1151,共7页Journal of Southeast University：Natural Science Edition

基　　金：国家重点研发计划资助项目(2020YFD1100201)。

摘　　要：针对DeepLabv3+对相似对象容易误判、小目标容易遗漏、预测输出存在空洞等问题,提出了一种融合通道注意力机制和空间注意力机制的FDA-DeepLab图像语义分割网络.首先,设计了一种结合通道注意力机制和空间注意力机制的特征融合模块,分别在4、8、16倍下采样特征图上使用该模块融合低层特征以弥补高层特征的不足;然后,针对训练样本的非均衡性问题,通过引入样本难度权重调节因子和类别权重改进了损失函数,提高了图像语义分割精度.最后,设计了消融和对比实验验证了所提网络.实验结果证明,该网络可有效提高模型的语义分割性能,在PASCAL VOC 2012验证集上相比原始模型MIoU值提高了1.2%,多尺度输入时MIoU值提高了1.9%.Aiming at the problems that DeepLabv3+ is easy to misjudge similar objects and miss small objects, and its prediction output is liable to have holes, a semantic segmentation network named fusion of dual attention DeepLab(FDA-DeepLab) incorporating channel attention mechanism and spatial attention mechanism was proposed. Firstly, a feature fusion module combining the channel attention mechanism and the spatial attention mechanism was designed, which was used to fuse low-level features to compensate for the lack of high-level features on 4, 8, and 16-fold downsampled feature maps, respectively. With this module, low-level features can be used to make up for the insufficiency of high-level features. Then, considering the sample imbalance problem, an improved focal loss function considering both the sample difficulty weight adjustment factor and class weight factor was proposed to improve the semantic segmentation performance. Finally, ablation and comparison experiments were designed to validate the proposed network. Experimental results show that the proposed FDA-DeepLab network can effectively improve the segmentation performance. Compared with the original model on the PASCAL VOC 2012 validation set, the network improves the mean intersection over union(MIoU) by 1.2%, and by 1.9% for multi-scale inputs.

关 键 词：注意力机制 语义分割 损失函数 DeepLabv3+ 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]