基于改进DeepLabv3+的草坪区域语义分割方法  

Semantic Segmentation Method of Lawn Area Based on Improved DeepLabv3+

作　　者：戚海洲 吴敬兵[1] 郭荣秋 QI Haizhou;WU Jingbing;GUO Rongqiu(School of Mechanical and Electrical Engineering,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学机电工程学院,武汉430070

出　　处：《自动化与仪表》2025年第4期132-136,141,共6页Automation & Instrumentation

摘　　要：针对智能割草机器人在运动过程中存在草地识别效果不理想的问题,提出了一种基于改进DeepLabv3+深度学习语义分割模型的智能割草机器人草坪识别方法。首先,该模型使用MobileNetv2代替Xception作为骨干网络进行初步特征提取来减少模型参数;其次,引入串行-并行空洞空间金字塔池化快速增加感受野以获取不同尺度的上下文信息;接着在解码部分添加CoordAttention注意力机制捕获跨通道的信息以及方向感知和位置感知的信息。实验结果表明,改进的语义分割模型在自制的草坪数据集上平均交并比达到了94.04%、平均像素精度达到了96.36%。与原始网络相比,模型参数量减少了48.0 M,推理时间下降了9.44 ms,在识别草坪环境中取得了更好的识别效果。In order to solve the problem that the grass recognition effect of the intelligent lawn mower robot is not ideal in the process of movement,a lawn recognition method of the intelligent lawn mower robot based on the improved DeepLabv3+deep learning semantic segmentation model was proposed.Firstly,the model uses MobileNetv2 instead of Xception as the backbone network for preliminary feature extraction to reduce the model parameters.Secondly,serial-parallel void space pyramid pooling was introduced to rapidly increase the receptive field to obtain contextual information at different scales.Next,CoordAttention is added to the decoding part to capture cross-channel information,as well as direction-aware and location-aware information.Experimental results show that the improved semantic segmentation model achieves an average intersection union ratio of 94.04%and an average pixel accuracy of 96.36%on the self-made lawn dataset.Compared with the original network,the number of model parameters is reduced by 48.0 M,and the inference time is reduced by 9.44 ms,which achieves better recognition effect in the recognition lawn environment.

关 键 词：草坪识别 深度学习 DeepLabv3+ 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]