Abstract: Accurate detection of immature and mature apples in orchards is crucial to the load management of early crops in orchards. This paper proposes a method that can detect apple maturity in real time and estimate the number of fruits in the whole fruit tree. In order to improve the detection ability of the YOLOXS network in complex scenes, Coordinate Attention (position attention) is added to the residual connection of the FPN (feature pyramid). In order to better detect the apple with dense growth, occluding and small size in the image, the position loss function is replaced from IoU_Loss to CIoU_Loss. The experimental results show that the mAP value of the improved YOLOXS detection algorithm proposed in this paper is increased by about 1.97% compared with the original algorithm, and the AP values of apples with low maturity, medium maturity and high maturity are 90.85%, 95.10% and 80.50%, respectively.

Key words: apple, YOLOXS, object detection, coordinate attention, maturity detection

摘要： 准确检测果园中未成熟与成熟的苹果对果园早期作物的负荷管理至关重要，提出一种能够实时检测苹果成熟度，并估算出整棵果树果实数量的方法。为提高YOLOXS网络在复杂场景下的检测能力，在FPN（特征金字塔）的残差连接处增加了Coordinate Attention（位置注意力）；为更好地检测图像中生长密集、存在遮挡、尺寸较小的苹果，将位置损失函数IoU_Loss更换为CIoU_Loss。试验结果表明，所提出的改进YOLOXS检测算法相较于原算法，mAP值提高约1.97%，苹果低成熟度、中等成熟度和高等成熟度的AP值分别为90.85%、95.10%和80.50%。

关键词: 苹果, YOLOXS, 目标检测, 位置注意力, 成熟度检测