多尺度特征融合的道路场景语义分割  

Semantic Segmentation for Road Scene Based on Multiscale Feature Fusion

作　　者：易清明[1] 张文婷 石敏[1] 沈佳林 骆爱文 Yi Qingming;Zhang Wenting;Shi Min;Shen Jialin;Luo Aiwen(College of Information Science and Technology,Jinan University,Guangzhou 510632,Guangdong,China)

机构地区：[1]暨南大学信息科学技术学院,广东广州510632

出　　处：《激光与光电子学进展》2023年第12期82-90,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(62002134);广东省基础与应用基础研究基金(2020A1515110645);广东省重点实验室项目(2021KSY001);羊城创新创业领军人才支持计划(2019019);暨南大学中央高校基本科研业务费项目(21620353)。

摘　　要：针对现有语义分割网络模型难以在参数量、推理速度和精确度中取得平衡的问题,设计了一种多尺度特征信息融合的轻量级网络模型(MIFNet)。MIFNet采用编码-解码结构,在编码部分利用分离策略和非对称卷积设计了轻量型特征提取瓶颈结构,且引入空间注意力机制与Laplace边缘检测算子组成边缘-空间融合模块,将空间信息和边缘信息进行融合得到丰富的特征信息。在解码部分引入通道注意力机制恢复特征图尺寸和细节信息完成语义分割。在Cityscapes和CamVid测试集上,MIFNet仅以0.82 M的参数量分别取得了73.1%和67.7%的分割精度,同时在单个GTX 1080Ti GPU下分别获得73.68 frame/s和85.16 frame/s的推理速度,表明该方法在参数量、推理速度和精确度3个指标上得到较好平衡,实现了轻量、快速、精准的语义分割。A lightweight network model based on multiscale feature information fusion(MIFNet)is developed in this study owing to the imbalance among the parameter amount,inference speed,and accuracy in many existing semantic segmentation network models.The MIFNet is constructed on the encoding-decoding architecture.In the encoding part,the split strategy and asymmetric convolution are flexibly applied to design lightweight bottleneck structure for feature extraction.The spatial attention mechanism and Laplace edge detection operator are introduced to fuse spatial and edge information to obtain rich feature information.In the decoding part,a new decoder is designed by introducing a channel attention mechanism to recover the size and detail information of the feature map for a complete semantic segmentation task.The MIFNet achieves accuracies of 73.1%and 67.7%on the Cityscapes and CamVid test sets,respectively,with only approximately 0.82 M parameters.Correspondingly,it reaches up to 73.68 frame/s and 85.16 frame/s inference speed,respectively using a single GTX 1080Ti GPU.The results show that the method achieves a good balance in terms of the parameter amount,inference speed,and accuracy,yielding a lightweight,fast,and accurate semantic segmentation.

关 键 词：图像处理 实时语义分割 Laplace边缘检测 注意力机制 多尺度特征信息融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]