【摘要】 针对现有道路场景语义分割算法存在准确性和实时性不兼容的问题，在DeepLabV3+基础上提出一种引入注意力机制的高效语义分割算法。提出一种并行主干特征提取网络来并行提取输入图像的语义信息和空间细节信息；改进通道域和空间域注意力机制模块并应用于主干特征提取网络之后；提出一个特征融合及上采样模块获取最终的图像分割结果。在Cityscapes数据集上验证所提算法的性能，结果表明：所提算法的平均交并比mIoU为74.54%,平均像素精度mPA为84.93%,处理一张图片的时间仅需45 ms;在模型分割精度和分割速度上达到更好的均衡，满足了自动驾驶系统对道路场景分割的要求。更多还原

【Abstract】 To address the incompatibility between accuracy and real-time performance of existing semantic segmentation algorithms for road scenes, an efficient semantic segmentation algorithm with attention mechanism is proposed based on DeepLabV3+. A parallel backbone feature extraction network is proposed to extract semantic and spatial detail information of input images in parallel. The channel domain and spatial domain attention mechanism modules are both improved and applied to the backbone feature extraction network. Finally, a feature fusion and upsampling module is proposed to obtain the final image segmentation results. The performance of the proposed algorithm is validated on the Cityscapes dataset. Results show the average intersection to union ratio(mIoU) of the algorithm is 74.54%, the average pixel accuracy(mPA) 84.93%, and the processing time for one image only 45 ms. The proposed algorithm better balances model segmentation accuracy and segmentation speed, meeting the requirements of the auto drive system for road scene segmentation.更多还原