引入反馈注意力的并行式多分辨率语义分割算法

Parallel Multi-resolution Semantic Segmentation Algorithm with Feedback Attention

中文关键词: 图像语义分割  反馈式注意力  多分辨率

英文关键词: image semantic segmentation  feedback attention  multi-resolution

中文摘要:

      目的 为了进一步提升语义分割精度，解决当前语义分割算法中特征图分辨率低下，低级信息特征随意丢弃，以及上下文重要信息不能顾及等问题，文中尝试提出一种融合反馈注意力模块的并行式多分辨率语义分割算法。方法 该算法提出一种并行式网络结构，在其中融合了高低分辨率信息，尽可能多地保留高维信息，减少低级信息要素的丢失，提升分割图像的分辨率。同时还在主干网络中嵌入了带反馈机制的感知注意力模块，从通道、空间、全局3个角度获得每个样本的权重信息，着重加强样本之间的特征重要性。在训练过程中，还使用了改进的损失函数，降低训练和优化难度。结果 经实验表明，文中的算法模型在PASCAL VOC2012、Camvid上的MIOU指标分别为77.78%、58.67%，在ADE20K上的也有42.52%，体现了出较好的分割性能。结论 文中的算法模型效果相较于之前的分割网络有一定程度的提升，算法中的部分模块嵌入别的主干网络依旧表现出较好的性能，展现了文中算法模型具备一定的有效性和泛化能力。

英文摘要:

      The work aims to propose a parallel multi-resolution semantic segmentation algorithm integrating feedback attention module, in order to further improve the accuracy of semantic segmentation and solve the problems of low resolution of feature map, random discarding of low-level information features and failure to take into account important contextual information in the current semantic segmentation algorithm. The algorithm exhibited a parallel network structure, which integrated high and low resolution information, retained high-dimensional information as much as possible, reduced the loss of low-level information elements, and improved the segmentation resolution. At the same time, a perceptual attention module with feedback mechanism was embedded in the backbone network to obtain the weight information of each sample from the perspectives of channel, space and global, focusing on strengthening the importance of characteristics among samples. In the training process, the improved loss function was also used to reduce the difficulty of training and optimization. Experiments showed that the proposed algorithm model achieved 77.78% and 58.67% MIOU indexes on Pascal voc2012 and Camvid respectively, and 42.52% on ADE20K, reflecting better segmentation performance. Compared with the previous segmentation network, the algorithm model has a certain degree of improvement. Some modules embedded in other backbone networks still show good performance, which shows that the algorithm model has certain effectiveness and generalization ability.