基于EDH-YOLO的轻量型温室番茄检测方法

Lightweight Greenhouse Tomato Detection Method Based on EDH−YOLO

摘要:

针对番茄采摘机器人识别算法包含复杂的网络结构和庞大的参数体量，严重限制检测模型的响应速度问题，本文提出一种改进的轻量级YOLO v5 （EDH-YOLO） 算法。为了能够在保持较高识别精度的同时大幅降低计算复杂度和模型内存占用量，引入EfficientNet-B0的轻量级网络作为YOLO v5算法的骨干网络;为了在训练过程中更好地定位目标物体的同时提高检测算法精度，引入DIoU损失函数;为了降低模型计算复杂度和提高模型表达能力，引入一种轻量化的Hardswish激活函数。实验结果显示，EDH-YOLO模型在识别效果损失较小的情况下，精确率、召回率和平均精度均值分别为95.9%、93.1%和96.8%，模型内存占用量仅为7.3 MB，检测速度为53.2 f/s，对比YOLO v5原模型内存占用量降低55.3%，EDH-YOLO模型检测速度提升60.0%。对比Faster R-CNN、YOLO v7和YOLO v8，EDH-YOLO模型在不同光照和遮挡等情况下具有较高鲁棒性。同时，将EDH-YOLO模型通过模型转换部署到安卓（Android）平台中，优化模型推理过程，满足温室复杂环境下番茄目标果实实时识别需求，可为设施环境下基于移动边缘计算的机器人目标识别及自动采收作业提供技术支持。

Abstract:

Considering the issue that the tomato-picking robot’s recognition algorithm has a complex network structure and a large number of parameters, which severely limit the detection model’s response speed, an improved lightweight YOLO v5（EDH ? YOLO） algorithm was proposed. To significantly reduce computational complexity and model size while maintaining high recognition accuracy, the lightweight EfficientNet?B0 network was introduced as the backbone of the YOLO v5 algorithm. To better locate target objects during training and improve detection accuracy, the DIoU loss function was introduced. To reduce the model’s computational complexity and enhance its expressive ability, the lightweight Hardswish activation function was introduced. Experimental results indicated that the EDH?YOLO model achieved accuracy, recall, and average precision of 95.9%，93.1%, and 96.8%，respectively, with minimal loss in recognition performance. The size of model was only 7.3 MB, and the detection speed reached 53.2 f/s. Compared with the original YOLO v5 model, the model size was reduced by 55.3%, and the detection speed of the EDH ? YOLO model was increased by 60.0%.Compared with Faster R?CNN, YOLO v7, and YOLO v8, the EDH?YOLO model demonstrated higher robustness under various lighting and occlusion conditions. Additionally, the EDH ?YOLO model was deployed on the Android platform through model conversion to optimize the inference process, meet real-time recognition requirements for tomato fruits in complex greenhouse environments, and provide technical support for robot target recognition and automatic harvesting operations based on mobile edge computing in facility environments.