Abstract: Aiming at the problem of fast recognition of ripe tomato fruits, an image dataset of ripe tomato fruit is collected and labeled for deep neural network model training. In addition, the target detection algorithm of the classic algorithm YOLOv3 is improved in light weight based on practical applications, so that it can be easily deployed on the embedded controller of tomatopicking robot. The activation function, clustering of anchor frame, nonmaximum suppression and loss function are also optimized, the efficiency and the stability of the algorithm is also improved. Through verification of the test set, the proposed improved YOLOv3 target detection algorithm can effectively recognize tomato ripe fruit in complex environment, including different density, different lighting conditions and even different occlusion degree. The final detection accuracy was 92.11%, recall rate was 86.21%, F1 score was 89%, mAP was 84.58%. Thus the experimental results demonstrate the feasibility, accuracy and robustness of the proposed method.

Key words: tomato, complex environment, deep neural network, YOLOv3 target detection, recognition of fruits

摘要： 针对西红柿成熟果实快速识别问题，采集并标注西红柿成熟果实的图像数据集，用于深度神经网络模型的训练，并基于实际应用对经典的YOLOv3目标检测算法进行模型轻量化改进，使其方便地部署到采摘机器人的嵌入式控制器上，同时对激活函数、锚框的聚类、非极大值抑制和损失函数等进行优化，提高算法运行的效率和稳定性。经测试集的验证，所提出的改进型YOLOv3目标检测算法在包括不同密集程度、不同光照条件和不同遮挡程度情况的复杂环境下最终检测精度为92.11%，召回率为86.21%，F1得分为89%，mAP为84.58%，即试验结果证明所提方法的可行性、准确性和鲁棒性。

关键词: 西红柿, 复杂环境, 深度神经网络, YOLOv3目标检测, 果实识别