Authors

Abstract

Thinning is an important agronomic process in pear production, thus the detection of pear inflorescence is an important technology for intelligentization of blossom thinning. In this paper, images of buds and flowers were collected under different natural conditions for model training, and the images were augmented by data augmentation methods. Model training was performed based on the YOLOv5s network with coordinate attention mechanism added to the backbone network and compared with the native YOLOv5s, YOLOv3, SSD 300, and Faster-RCNN algorithms. The mAP, F1 score and recall of the algorithm reached 93.32%, 91.10%, and 91.99%. The model size only took up 14.1 MB, and the average detection time was 27 ms, which are suitable for application in actual intelligent blossom thinning equipment.

Abstract in Chinese

疏花疏果是梨生产中一项重要的农艺环节，而梨树花序识别是疏花智能化过程中的一项重要技术。本文在不同的自然条件下采集了梨树花苞与花朵的图像，通过数据增强方法对图像进行了扩充。在YOLOv5s主干网络中增加CA注意力机制，将其与原生YOLOv5s、YOLOv3、SSD 300、Faster-RCNN算法进行对比。结果表明，改进的YOLOv5s-CA模型优于其他模型，其mAP、F1得分、Recall分别达到了93.32%、91.10%和91.99%，模型大小仅为14.1 MB，平均检测时间27 ms，更适用于智能化疏花设备中。