自然环境下黄绿柑橘检测通用模型的构建

Construction of a general model for yellow-green citrus detection in natural environment

柑橘存在黄色和绿色两种颜色特征,自然环境下柑橘目标检测困难。深度学习在目标检测领域已经实现了实时检测,基于深度学习进行柑橘检测,探索了两种颜色特征的柑橘检测通用模型。建立了自然环境下柑橘的图像数据集,采用3种深度学习检测模型:基于VGG16的Faster R-CNN、基于Resnet的Faster R-CNN以及YOLOv5s,分别对自然环境下的黄色和绿色柑橘进行对比试验。试验表明:对于黄色柑橘,YOLOv5s模型检测方法在柑橘测试集上的精确率、召回率、F1以及平均精度(AP)分别为91.90%、99.00%、0.94和97.40%,平均检测速度为27帧/s;对于绿色柑橘,YOLOv5s模型检测方法在柑橘测试集上的精确率、召回率、F1以及AP分别为96.50%、98.00%、0.96和97.20%,平均检测速度为32帧/s。YOLOv5s与基于VGG16的Faster R-CNN模型及基于Resnet的Faster R-CNN模型相比:对于黄色柑橘,平均精度分别提高了45.51%和41.18%,平均检测速度分别提高了22和23帧/s;对于绿色柑橘,平均精度分别提高了4.38%和4.13%,平均检测速度分别提高了25和26帧/s。结果表明,对于两种颜色特征的柑橘,YOLOv5s检测模型通用性更好,检测速度更快,更适用于柑橘检测研究。

In response to the national smart agriculture and rural revitalization strategy,the application of deep learning technology to forestry and agriculture has important practical significance for improving agricultural productivity,big data management under smart agriculture,and helping rural revitalization.Citrus is a common fruit in the south of China.In natural environments,citrus has two color characteristics,namely yellow and green.In addition,the complex orchard environment makes it difficult to detect citrus targets.In recent years,deep learning has achieved real-time detection in the field of target detection,applying deep learning to citrus detection,exploring a general model of citrus detection with two color features,and providing technical supports for the research on citrus visual detection in natural environments.In this study,an image data set of citrus in the natural environments was established,and three deep learning detection models were used,which were Faster R-CNN based on VGG16,Faster R-CNN based on Resnet and YOLOv5s,respectively,for comparing the yellow and green citrus in the natural environments.The experiment results showed that,for yellow citrus,the precision,recall rate,F1 and AP of the YOLOv5s model detection method on the citrus test set were 91.90%,99.00%,0.94 and 97.40%,respectively,and the average detection speed was 27 frames/s;for green citrus,the precision,recall rate,F1 and AP of the YOLOv5s model detection method on the citrus test set were 96.50%,98.00%,0.96,and 97.20%,respectively,and the average detection speed was 32 frames/s.For two colors of citrus,each evaluation index of the YOLOv5s model had a good performance.Compared with the VGG16-based Faster R-CNN model and the Resnet-based Faster R-CNN model,for yellow citrus,the precisions of YOLOv5s were increased by 51.54%and 40.84%,the recall rates were increased by 41.08%and 37.83%,the average precisions were increased by 45.51%and 41.18%,and the average detection speeds were increased by 22 and 23 frames/s,respectively;for green citrus,the precisions were increased by 21.38%and 17.88%,the recall rates were increased by 3.65%and 3.36%,the average precisions were improved by 4.38%and 4.13%,and the average detection speeds were improved by 25 and 26 frames/s,respectively.The results showed that for citrus with two color characteristics,the performance of YOLOv5s detection model was the best,indicating that YOLOv5s had better versatility and faster detection speed,and was more suitable for the citrus detection.