基于轻量化改进DeepLabV3+的果园道路语义分割方法

Orchard Road Semantic Segmentation Based on Lightweight Improved DeepLabV3+

摘要:

针对果园行间道路光影紊乱、纹理模糊、遮挡频繁等现象,导致现有语义分割方法精度低、速度慢与模型轻量化难等问题,本文提出一种基于改进DeepLabV3+的果园道路轻量化分割方法。采用轻量化MobileNetV4ConvSmall作为主干网络,提升模型推理效率;引入ECA通道注意力机制,增强模型对道路关键特征的鉴别能力;设计Lite-ASPP轻量化多尺度上下文模块,融合不同感受野特征,提升对道路边界与细长结构建模性能。自建果园道路数据集上试验结果表明,改进模型平均交并比与平均像素精度分别达91.07%和95.96%,平均推理帧率为511.08 f/s,参数量仅为2.62×10^6,相较于PSPNet、HRNet、SegFormer及DeepLabV3+等模型,本文方法在分割精度与实时性上均取得较优结果,在不同地表状态与干扰环境下均能实现连续、稳定的道路区域分割。研究结果可为果园作业装备实时视觉导航提供理论依据与技术参考。

Abstract:

Orchard inter-row roads are commonly affected by uneven illumination, blurred ground textures, frequent occlusions, and complex-background interference, which reduce the segmentation accuracy, inference speed, and lightweight deployment capability of existing semantic segmentation methods. To solve these problems, a lightweight orchard road semantic segmentation method was proposed based on an improved DeepLabV3+ network, aiming to improve the operational efficiency of agricultural machinery in orchard inter-row environments. Firstly, the lightweight MobileNetV4ConvSmall network was adopted as the backbone to replace the original feature extraction network, thereby reducing computational complexity and significantly improving model inference efficiency. Secondly, an efficient channel attention (ECA) mechanism was introduced after the high-level features of the backbone network to enhance the discriminative representation of key road features under complex lighting and occlusion conditions. Furthermore, a Lite-ASPP lightweight multi-scale context module was designed to integrate features with different receptive fields, improving the representation capability for road boundaries, slender structures, and weak edge regions. Experimental evaluation was conducted on a self-constructed orchard road dataset. The results showed that the improved model achieved a mean intersection over union (mIoU) of 91.07% and a mean pixel accuracy (mPA) of 95.96%, with an average inference speed of 511.08 f/s and only 2.62×10^6 parameters. Compared with PSPNet, HRNet, SegFormer, and the DeepLabV3+, the proposed method achieved better overall performance in segmentation accuracy and real-time inference. Under different ground conditions and environmental disturbances, the model can continuously and stably extract road regions, providing theoretical support and technical reference for real-time visual navigation of orchard operation equipment in practice.