Abstract: Fruit recognition is an important part of visual detection technology, and its recognition accuracy is easily affected by the complex growth environment and fruit state. This paper takes tomato fruits in six complex growth states, such as single, cluster, illumination, shadow, occlusion and overlapping in greenhouse environment as objects, and proposes a tomato fruit recognition method based on the combination of improved YOLOv4 network model and transfer learning. Firstly, the ImageNet dataset and the 16 convolutional layers at the front end of the VGG network model were used for model parameters pretraining, and the trained model parameters were initialized to improve the weights of the model to replace the original initialization operation. Then the tomato dataset was used for training in the new model combined with the convolutional layers of VGG19 and the backbone network of YOLOV4. The optimal weight was obtained to detect tomato fruits in complex environment. Finally, the improved model was compared with Faster RCNN, YOLOv4-Tiny, and YOLOv4 network models. The results showed that the average detection accuracy (mAP) of the improved model under six complex environments for tomato fruits was 89.07%, 92.82%, 92.48%, 93.39%, 93.20%, and 93.11%, and the F1 scores at different maturity levels of ripe, halfripe, and immature were 84%, 77%, and 85%, respectively, and their recognition accuracy is better than that of the comparison models. The method in this paper realizes effective detection and recognition of tomato fruits in six complex environments, which provides a theoretical basis for the intelligent picking of tomato fruits.

Key words: tomato, complex environment, fruit detection, network model, YOLOv4, transfer learning

摘要： 果实识别是视觉检测技术重要的环节，其识别精度易受复杂的生长环境及果实状态的影响。以大棚环境下单个、一簇、光照、阴影、遮挡、重叠6种复杂生长状态下的番茄果实为对象，提出一种基于改进YOLOv4网络模型与迁移学习相结合的番茄果实识别方法。首先利用ImageNet数据集与VGG网络模型前端16卷积层进行模型参数预训练，将训练的模型参数初始化改进模型的权值以代替原始的初始化操作，然后使用番茄数据集在VGG19的卷积层与YOLOV4的主干网络相结合的新模型中进行训练，获得最优权重实现对复杂环境下的番茄果实进行检测。最后，将改进模型与Faster RCNN、YOLOv4-Tiny、YOLOv4网络模型进行比较。研究结果表明，改进模型在6种复杂环境下番茄果实平均检测精度值mAP达到89.07%、92.82%、92.48%、93.39%、93.20%、93.11%，在成熟、半成熟、未成熟3种不同成熟度下的F1分数值为84%、77%、85%，其识别精度优于比较模型。本文方法实现了在6种复杂环境下有效地番茄果实检测识别，为番茄果实的智能采摘提供理论基础。

关键词: 番茄, 复杂环境, 果实检测, 网络模型, YOLOv4, 迁移学习