基于改进YOLO v8n网络的番茄成熟度实时检测算法

Improved YOLO v8n Network for Real-time Detection of Tomato Maturity

摘要:

为应对番茄采摘面临的果农老龄化、劳动力短缺和人工成本上涨等挑战，解决在复杂果园环境下番茄采摘机器人视觉系统成熟度检测精度低和实例分割不准确等问题，本文提出一种基于改进YOLO v8n网络的番茄成熟度实时检测算法。首先，通过在YOLO v8n网络中引入通道嵌入位置注意力模块和改进大核卷积块注意力模块，能够在浅层网络保留番茄目标位置信息，建立目标区域之间的长距离依赖关系，从而增加YOLO v8n网络对显著番茄特征的关注。然后，在LaboroTomato数据集上进行了对比实验，改进YOLO v8n相较于原YOLO v8n网络，检测和分割的mAP@50和mAP@50-95分别提高0.4、1.4个百分点和0.3、1.2个百分点。最后，实现了改进YOLO v8n网络在低成本、低算力和低功耗的Jetson Nano平台上的轻量化部署，模型内存占用量由满溢减少到2.4 GB，推理速度加倍。该研究可为番茄采摘机器人在复杂场景下实时、准确检测番茄成熟度提供技术支撑。

Abstract:

To address the numerous challenges faced in tomato harvesting, such as the aging of farmers, labor shortages, and rising labor costs, and resolve issues related to the low maturity detection accuracy and inaccurate instance segmentation of tomato harvesting robots in complex orchard environments, an improved YOLO v8 network-based real-time tomato maturity detection algorithm was proposed. Firstly, by introducing the channel embedded positional attention module and an improved large kernel convolutional block attention module into the YOLO v8n network, the algorithm can effectively retain the positional information of tomato targets in the shallow network layers and establish long-range dependencies between target regions, thereby significantly increasing the attention of the YOLO v8n network to critical tomato features. Then a series of comprehensive and rigorous comparative experiments were conducted on the LaboroTomato dataset, demonstrating that the improved YOLO v8n network achieved 0.4 percentage points, 1.4 percentage points, and 0.3 percentage points, 1.2 percentage points improvements in detection and segmentation mAP@50 and mAP@50-95, respectively, compared with that of the original YOLO v8n network. Finally, the improved YOLO v8n network was lightweight deployed on the low-cost, low-computation, and low-power Jetson Nano platform, successfully reducing memory usage from overflow to 2.4 GB and doubling the inference speed. The research result can provide robust technical support for the real-time and accurate detection of tomato maturity by tomato harvesting robots in complex scenarios, significantly enhancing the overall efficiency and effectiveness of automated tomato harvesting operations.