目标检测算法Yolov8用于转色柑桔果实检测的改进

Improvement of Yolov8 for detection of citrus fruit color-changing

为提高在自然环境下对柑桔果实的识别,针对当前柑桔目标检测中树冠大、果实个体小、密集且遮挡严重等导致果实检测难度大的问题,基于自然状态下转色中后期温州蜜柑单侧完整树冠图像构建的果实数据集,提出了一种在Yolov8检测模型上添加小目标检测层的I-Yolov8检测模型。结果表明,以自然环境下的冠层为背景,丰富了数据集的目标特征,而添加的小目标检测层可用于检测4像素×4像素以上的目标。模型的训练平均精度(mAP)达到93.5%,相比Yolov8提升了1.3百分点。在晴天和阴天两个自然场景下分别进行预测,I-Yolov8和Yolov8的检测精确率均为100%;I-Yolov8的召回率分别达72.45%和91.61%,相比Yolov8分别提升了16.33和14.63百分点。I-Yolov8网络模型对于自然环境中柑桔的检测精度高,具备较高的应用潜力。

In order to improve the recognition of citrus fruit in natural environment,aiming at the problem that fruit detection is difficult due to the large crown,small individual fruit,dense fruit and serious occlusion in the current citrus target detection,an I-Yolov8 detection model with small target detection layer added to the Yolov8 detection model was proposed based on the fruit dataset constructed by the unilateral complete crown image of Citrus unshiu Marc.cv.Miyagawa wase during mid-late stages of fruit color-changing period under natural conditions.The results showed that the canopy in the natural environment enriches the target features of the dataset,and the added small target detection layer could be used to detect the targets with or above 4 pixel×4 pixel,and the mean Average Precision(mAP)of the model could reach 93.5%,which was 1.3%higher than that of Yolov8.The detection accuracy of I-Yolov8 and Yolov8 was 100%,respectively in sunny and cloudy natural scenarios,and the recall rate of I-Yolov8 was 72.45%and 91.61%,respectively,which was 16.33 and 14.63 percent higher than that of Yolov8,respectively.Therefore,the I-Yolov8 network model has high detection accuracy for citrus fruit in the natural environment and high application potential.