基于YOLOv5s的深度学习在自然场景苹果花朵检测中的应用

疏花是苹果栽培的重要管理措施，机械疏花是目前最具有发展潜力的疏花方式，花朵的高效检测是疏花机器人高效作业的重要保障。该研究基于机器视觉与深度学习技术，提出了一种基于YOLOv5s深度学习的苹果花朵检测方法，在对田间拍摄得到的苹果花朵图像标注后，将其送入微调的YOLOv5s目标检测网络进行苹果花朵的检测。经测试，模型的精确率为87.70%，召回率为0.94，均值平均精度(mean Average Precision, mAP)为97.20%，模型大小为14.09 MB，检测速度为60.17 帧/s，与YOLOv4、SSD和Faster-RCNN模型相比，召回率分别提高了0.07、0.15、0.07，mAP分别提高了8.15、9.75和9.68个百分点，模型大小减小了94.23%、84.54%、86.97%，检测速度提升了126.71%、32.30%、311.28%。同时，该研究对不同天气、颜色和光照情况下的苹果花朵进行检测，结果表明，该模型对晴天、多云、阴天、小雨天气下苹果花朵的检测精确率分别为86.20%、87.00%、87.90%、86.80%，召回率分别为0.93、0.94、0.94、0.94，mAP分别为97.50%、97.30%、96.80%、97.60%。该模型检测白色、粉色、玫红色和红色花朵的精确率分别为84.70%、91.70%、89.40%、86.90%，召回率分别为0.93、0.94、0.93、0.93，mAP分别为96.40%、97.70%、96.50%、97.90%。该模型检测顺光和逆光条件下苹果花朵的精确率分别为88.20%、86.40%，召回率分别为0.94、0.93，mAP分别为97.40%、97.10%。结果表明YOLOv5s可以准确快速地实现苹果花朵的检测，模型具有较高的鲁棒性，且模型较小，更有利于模型的迁移应用，可为疏花器械的发展提供一定的技术支持。

Application of deep learning using YOLOv5s to apple flower detection in natural scenes

Flower thinning is one of the most important management measures in apple cultivation. Mechanical thinning has been the most promising way for thinning flowers in recent years. Accurate and rapid detection of flowers can be critical to the highly efficient operation of flower thinning robots. In this study, an apple flower detection was proposed using machine vision and YOLOv5s deep learning. 3 005 apple flower images were collected, including 1 611 apple images on sunny days, 512 on cloudy, 519 on overcast sky days, and 363 on light rainy days. Two lighting conditions were considered, where 1 830 apple images under the front lighting and 1 175 apple images under the backlight. Two occlusion situations were selected, where 1 602 apple images with occlusion, and 1 403 apple images without occlusion. The apple flower images were taken to annotate in the field, and then sent to the fine-tuned YOLOv5s target detection network for the detection of the apple flower. 300 iterations of training were implemented after the test. The better performance was achieved, where the precision of the model was 87.70%, the recall was 0.94, the mean average precision was 97.20%, the model size was 14.09 MB, and the detection speed was 60.17 f/s. Specifically, the recall increased by 7, 15, and 7 percental points, respectively, compared with the YOLOv4, SSD, and Faster-RCNN models, while the mAP increased by 8.15, 9.75, and 9.68 percental points, respectively, the model size decreased by 94.23%, 84.54%, and 86.97%, respectively, as well as the detection speed increased by 126.71%, 32.30%, and 311.28%, respectively. At the same time, the study detected apple flowers in different weather, colors and light conditions. The results showed that the precision values of the model to detect the white, pink, rose and red flowers were 84.70%, 91.70%, 89.40%, and 86.90%, respectively, while the recall were 0.93, 0.94, 0.93, and 0.93, respectively, as well as the mean average precision were 96.40%, 97.70%, 96.50%, and 97.90%, respectively. The precision values of the model for detecting apple blossoms under sunny, cloudy, overcast, and light rain were 86.20%, 87.00%, 87.90%, and 86.80%, respectively, the recall were 0.93, 0.94, 0.94, and 0.94, respectively, as well as the mean average accuracy were 97.50%, 97.30%, 96.80%, and 97.60%, respectively. The precision values of this model for detecting apple flowers under forwarding and backlight conditions were 88.20% and 86.40%, respectively, the recalls were 0.94 and 0.93, respectively, as well as the mean average accuracy was 97.40% and 97.10%, respectively. Consequently, the YOLOv5s can be expected to detect the apple flowers accurately and rapidly. The higher robustness and the smaller size were more conducive to the migration and application of the model. The finding can provide strong technical support to develop the flower thinning equipment.