| 摘 要: 针对实时语义分割中高精度与低延迟难以兼顾的问题,提出双分支多尺度特征复用网络(TMMNet)。该网络采用语义分支与细节分支协同设计,分别提取深层语义与边缘细节信息,并通过多尺度信息提取融合模块(MEF)与多尺度信息融合还原模块(MFR)实现跨尺度特征的高效整合。在 Cityscapes和 Camvid数据集上的实验结果表明,TMMNet的 MIoU 分别为74.9%和67.8%,同时推理速度高达94frame/s及128frame/s,明显优于对比方法。综上,该方法在精度、速度与模型复杂度间实现了理想平衡,为自动驾驶、机器人导航等实时视觉应用提供了有力的技术支撑。 |
| 关键词: 自动驾驶 实时语义分割 特征复用 多尺度信息 |
|
中图分类号: TP391
文献标识码: A
|
|
| Real-Time Semantic Segmentation Method Based on Dual-Branch Multi-Scale Feature Reuse Network |
|
MIAO Deting1, FANG Yinfeng1, ZHANG Xuguang2
|
(1.College of Communication Engineering, Hangzhou Dianzi University, Hangzhou 310018, China;
2.School of Media Engineering, Communication University of Zhejiang, Hangzhou 310018, China)
221080054@hdu.edu.cn; yinfeng.fang@hdu.edu.cn; 20250002@cuz.edu.cn
|
| Abstract: To address the challenge of balancing high accuracy and low latency in rea-l time semantic segmentation, a Twin-Branch Mult-i Scale Feature Mux Network,TMMNet ( TMMNet) is proposed. This network employs a collaborative design of a semantic branch and a detail branch to extract deep semantic and edge detail information, respectively. It achieves efficient integration of cross-scale features through a Mult-i scale Extraction Fusion module (MEF) and a Mult-i scale Fusion Restoration module (MFR). Experimental results on the Cityscapes and Camvid datasets demonstrate that TMMNet achieves MIoU scores of 74.9% and 67.8% , respectively, with inference speeds reaching 94 frame/s and 128 frame/s, significantly outperforming comparison methods. In conclusion, the method achieves an ideal balance among accuracy, speed, and model complexity, providing powerful technical support for rea-l time vision applications such as autonomous driving and robot navigation. |
| Keywords: autonomous driving rea-l time semantic segmentation feature reuse mult-i scale information |
