基于混合注意力的实时语义分割算法  

Attention Fusion for Real-Time Semantic Segmentation

作　　者：严广宇 刘正熙[1] YAN Guang-yu;LIU Zheng-xi(College of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《现代计算机》2020年第10期34-38,共5页Modern Computer

摘　　要：实时语义图像分割算法出于模型推理速度考虑,通常会采取减少网络深度、宽度,缩小输入图片分辨率等方式来限制模型计算量和参数量。但上述这些操作会导致模型分割精度损失过多。针对此问题,提出使用混合注意力机制,在不引入过多参数和降低推理速度的情况下,使用空间注意力与通道注意力分支分别基于编码器特征图学习自适应权重,然后与特征图进行混合叠加。实验结果表明,在自行设计的网络中,综合运用空间注意力和通道注意力可以使模型在Cityscapes验证集上的mIOU提高1.82%。对于1024×512的图像,在NVIDIA GTX 1080ti显卡上模型推理速度为58fps。实验结果基本验证提出方法的有效性。Real-time semantic image segmentation is based on the speed of model inference.Generally,the amount of model calculation and parame ters are limited by reducing the depth or width of the network or reducing the resolution of the input images.However,these operations will cause excessive loss of model segmentation accuracy.In view of this problem,we propose to use mixed attention,which uses spatial atten tion and channel attention branch to learn adaptive weights based on encoder feature maps without adding too many parameters and reduc ing inference speed.After that,we use residual branches to fuse the attention features with the encoder feature maps.Experiments show that,in our self-designed network,the comprehensive use of spatial attention and channel attention can improve the model's mIOU on the Cityscapes validation set by 1.82%.For 1024×512 images,the model inference speed is 58 FPS on GTX 1080ti.The experimental results basically verify the effectiveness of our proposed method.

关 键 词：实时 语义分割 空间注意力 通道注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]