基于改进DeepLabV3+的丘陵田间道路图像分割方法研究  

Research on Hilly Field Road Image Segmentation Method Based on Improved DeepLabV3+

作　　者：李法霖 石军锋[1] 梁新成[1] 李云伍[1] 刘鹏 陈欣 LI Falin;SHI Junfeng;LIANG Xincheng;LI Yunwu;LIU Peng;CHEN Xin(College of Engineering and Technology,Southwest University,Chongqing 400715,China)

机构地区：[1]西南大学工程技术学院,重庆400715

出　　处：《西南大学学报（自然科学版）》2024年第8期172-183,共12页Journal of Southwest University(Natural Science Edition)

基　　金：重庆市科技局项目(cstc2021jcyj-msxmX1062);贵州省科技计划项目(黔科合支撑[2022]一般168).

摘　　要：为解决丘陵地区智能农机装备因道路狭窄、路况复杂导致道路信息提取精度低和推理速度慢等问题,以丘陵田间道路作为研究对象制作数据集,提出一种基于改进DeepLabV3+的丘陵田间道路图像分割方法.首先在编码器模块中使用轻量化的主干网络G_Ghost_RegNetX_4.0GF提取图像特征,保证精度并减小模型参数数量.再采用轻量级的空洞空间金字塔池化模块,将不同尺度特征融合.试验结果表明,改进模型的平均交并比和推理速度分别为87.6%及116.08 f/s,与当前主流图像分割网络FCN、DeepLabV3及PSPNet相比,MIoU分别提升了0.8%,2.2%,1%,推理速度分别为对比网络的1.33,1.83,1.76倍.所提模型的参数总量为14.41×10^(6),浮点计算量为49.34×10^(9),模型参数及计算量大幅减小.改进后的算法具有较高的检测精度和推理速度,有利于解决智能农机装备在丘陵田间道路上行驶的自主导航问题.To tackle the challenges of low precision in road information extraction and slow inference speeds associated with narrow roads and complex terrain in hilly areas,it is significant to initiate the creation of a dataset,with a specific focus on hilly field roads as the research subject.It is also valuable to introduce a hilly field road image segmentation method,leveraging an enhanced DeepLabV3+model.In the encoder module,it is crucial to integrate a lightweight backbone network,G_Ghost_RegNetX_4.0GF,which facilitated precise feature extraction while simultaneously reducing the model s parameter count.Additionally,it is necessary to incorporate the Lite-RASPP module to fuse features of varying scales.Relevant experimental findings underscore the success of this enhanced model,achieving a remarkable Mean Intersection over Union(MIoU)of 87.6%and an impressive inference speed of 116.08 f/s.When compared to prevailing image segmentation networks like FCN,DeepLabV3,and PSPNet,our model exhibited a substantial MIoU increase of 0.8%,2.2%,and 1%respectively,while significantly outpacing them in terms of inference speed,being 1.33,1.83,and 1.76 times faster,respectively.Furthermore,our proposed model boasted a lean parameter count of 14.41×10^(6)and a floating-point computation amount of 49.34×10^(9),substantially reducing both the model s parameters and computational demands.This optimized field image segmentation algorithm not only demonstrated enhanced detection accuracy but also delivered a superior inference speed,holding significant promise for facilitating the autonomous navigation of intelligent agricultural apparatus on hilly field roads.

关 键 词：丘陵道路 机器视觉 场景识别 语义分割 神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]