Abstract: The artificial harvesting of Chinese chestnut is low efficiency, high labor intensity and easy to hurt people. Chestnut positioning and picking by unmanned aerial vehicle (UAV) based on machine vision are both efficient and safe. To quickly identify and accurately locate chestnut targets in natural environment, a YOLOvc chestnut target detection method with improved YOLOv3 network structure is proposed. By adding the CBAM attention mechanism module integrating channel attention and spatial attention mechanism to the front end of the network YOLOv3 layer, the extracting little target features ability of deep learning network model is improved. Secondly, the focal loss function is added on the basis of the original loss function of YOLOv3 to improve the detection and identification ability of difficult samples such as chestnut occlusion. The results show that YOLOvc algorithm can effectively detect chestnut, its accuracy rate and average accuracy are 89.35% and 89.37% respectively.The results of the ablation experiments showed that the precision of improved YOLOv3 convolutional neural network was 2.21% higher than the YOLOv3. The research results show that the deep learning algorithm YOLOvc by adding attention mechanism and focus loss function to YOLOv3 can effectively realize chestnut detection and localization on trees and provide effective technical support for chestnut harvesting by UAV. 

Key words: chestnut peng, YOLOv3, object detection, CBAM, focal loss

摘要： 板栗人工采收效率低、劳动强度大、易伤人,基于机器视觉的无人机采摘板栗栗蓬既高效又安全、为了快速识别和准确定位自然环境下板栗栗蓬目标,提出一种改进YOLOv3网络结构的YOLOvc栗蓬目标检测方法。通过在网络YOLOv3层前端添加融合通道注意力与空间注意力机制的CBAM注意力机制模块,提高深度学习网络模型对小栗蓬目标特征提取能力。在YOLOv3原有损失函数基础上添加焦点损失函数,提高对栗蓬遮挡等难分样本的检测识别能力。试验结果表明，YOLOvc算法能够有效检测板栗栗蓬，其查准率和平均精度分别达到89.35%、89.37%。消融试验对比结果显示,改进YOLOv3卷积神经网络对板栗树上栗蓬的检测识别精度比YOLOv3提高2.21%。研究结果表明，对YOLOv3添加注意力机制和焦点损失函数的深度学习算法YOLOvc可有效实现板栗树上栗蓬检测定位，为无人机采收板栗提供有效技术支持。

关键词: 板栗栗蓬, YOLOv3, 目标检测, CBAM, 焦点损失