作者:王小玉，李志斌
Authors:WANG Xiao yu,LI Zhi bin摘要:针对实时语义分割任务中需要同时兼顾位置信息和语义信息的问题 ,提出一种改进特征融合的实时语义分割方法 。该方法由卷积神经网络、轻量级注意力模块(light attention module, LAM) 和双通道特征融合模块 (bilateral feature fusion module, BFFM)组成 。首先 ,使用卷积神经网络结合轻量级注意力模块快速提取图像的位置信息和语义信息 。然后 ,使用双通道特征融合模块指导位置信息和语义信息的特征图融合 。所提方法在 CamVid上 ,平均交并比达到 67. 8% ,分割速度可达到 52. 6 帧/s。在 Cityscapes 上 ,平均交并比达到 73. 5% ,分割速度可达到 31. 8 帧/s。实验结果表明,提出的分割方法满足分割的准确性和实时性要求 ,能够适用于实时语义分割任务中。
Abstract:Aiming at the problem that both location information and semantic information need to be considered in real-time semantic segmentation tasks, we proposed a real-time semantic segmentation method based on improved feature fusion. The method consists of a convolution neural network, a light attention module (light attention module, LAM) , and a bilateral feature fusion module (bilateral feature fusion module, BFFM). Firstly, quickly extract the location information and semantic information of the image by combining the convolutional neural network with the lightweight attention module. Then, the bilateral feature fusion module is used to guide the feature fusion of location information and semantic information. The results of the method on the data set of CamVid, mIoU reached 67. 8% , and the running speed reached 52. 6 fps. On the data set of Cityscapes, mIoU reached 73. 5% , and the running speed reached 31. 8 fps. The results show that the proposed segmentation method meets the accuracy and real-time requirements of segmentation, and can be applied to real-time semantic segmentation tasks.

PDF全文下载地址:

可免费Download/下载PDF全文