基于改进YOLOv5s的自然环境下番茄成熟度检测方法

Tomato ripening detection in natural environment based on improved YOLOv5s

【目的】在番茄识别任务中,现有的目标识别算法速度慢、对遮挡番茄以及小番茄识别准确率低,影响了其在嵌入式设备上的部署和应用。为实现复杂环境下农业机器人对番茄果实的快速准确识别,该研究提出一种基于改进的YOLOv5s模型的番茄成熟度识别方法。【方法】结合番茄生长环境的分布特点,引入MobileNetv3网络和ECANet注意机制,改进YOLOv5s目标检测算法。【结果】改进后的YOLOv5s(Im-YOLOv5s)与YOLOv5相比,准确率、召回率和平均准确率分别提高3.4%、2.4%和2.3%,权重大小降低了48.6%,检测速度提高了52.9%,提高了检测性能,缩短了模型推理时间。【结论】与多种主流目标检测模型相比,改进后的YOLOv5s对番茄的成熟度的漏检和误检大大减少,识别效果更好,具有良好的鲁棒性和实时性,满足对不同成熟度番茄的精准实时识别需求,适合在嵌入式设备上的部署和应用,可为番茄自动化采摘提供技术支持。

[Objective]In the tomato recognition task,the existing target recognition algorithms are slow and have low accuracy in recognizing occluded tomatoes and small tomatoes,which affects their deployment and application on embedded devices.In order to realize the fast and accurate recognition of tomato fruits by agricultural robots in complex environments,this study proposes a tomato ripeness recognition method based on the improved YOLOv5s model.[Method]Combined with the distribution characteristics of tomato growing environment,MobileNetv3 network and ECANet(Efficient Channel Attention Network)mechanism were introduced to improve the YOLOv5s target detection algorithm.[Result]Compared with YOLOv5,the results of the improved YOLOv5s(Im-YOLOv5s)showed that the accuracy,recall,and average accuracy were improved by 3.4%,2.4%,and 2.3%,the weight size was reduced by 48.6%,and the detection speed was increased by 52.9%,which improved the detection performance and shortened the model inference time.[Conclusion]Compared with a variety of mainstream target detection models,the improved YOLOv5s algorithm greatly reduces the omission and false detection of tomato ripeness.With better recognition effect,good robustness and real-time performance,it meets the demand for accurate real-time identification of tomatoes with different ripeness,and it is suitable for the deployment and application on embedded devices,thus providing technical support for automated tomato picking.