条带池化注意力的实时语义分割算法

Stripe Pooling Attention for Real-Time Semantic Segmentation

针对目前面向注意力机制语义分割算法不能兼顾分割精度与实时性,以及难以满足在现实场景中应用的问题,提出一种基于条带池化注意力的实时语义分割算法.首先采用轻量级骨干网络提取特征信息,结合不同网络层语义差异构建特征融合模块获得不同尺度的上下文信息以提升分割精度;然后利用基于注意力的条带注意力模块(SAM)提高远距离信息被弱化的注意力,并在SAM中加入水平方向的条带池化以降低编码全局上下文的运算量.实验结果表明,所提算法能够得到较高分割精度且满足实时性要求;在Cityscapes测试集上平均交并比为70.6%,分割速度达到了92帧/s;在CamVid测试集上平均交并比为66.4%,分割速度达到了196帧/s.

In order to solve the problem that it is difficult to meet the application in real scene because of the attention mechanism semantic segmentation algorithm cannot achieves a good balance between segmenta-tion speed and accuracy.We proposed a lightweight real-time semantic segmentation algorithm based on strip-pooling attention.Firstly,lightweight backbone network was used to extract feature information,and a feature fusion module was constructed to obtain context information at different scales to improve the seg-mentation accuracy.Secondly the attention-based strip attention module(SAM)is used to improve the atten-tiveness of remote information,and horizontal strip pooling is added to SAM to reduce the computation of encoding global context.Experimental results show that the proposed algorithm can achieve high segmenta-tion accuracy and meet the reai-time requirements,mIoU reached 70.6%on Cityscapes test set and the av-erage segmentation speed is 92 frames per second;mIoU reached 66.4 on CamVid test set and the average segmentation speed is 196 frames per second.