基于颜色特征量化和改进YOLO v8的番茄成熟度分级检测方法

Tomato Ripeness Grading Detection Method Based on Color Feature Quantification and Improved YOLO v8

摘要:

番茄的成熟度与其品质密切相关，是生产中采摘和分拣等环节的重要依据。针对作物成熟度分级检测系统功能简单，人工升级系统成本较大的问题，本文以番茄为例，采集并构建自然场景下番茄图像数据集，设计以番茄果实成熟度分级算法为基础的番茄图像半自动标注算法对采集后的数据进行标注，在YOLO v8模型基础上，将FPN结构替换为BiFPN结构实现更高效的多尺度特征融合，利用SE注意力机制对空间和通道进行融合特征提取，引入Focal SIoU损失函数对预测框与真实框之间的角度差异进行度量，构建基于颜色特征量化和改进YOLO v8的番茄成熟度分级检测模型YOLO v8_BFS，识别番茄生长过程的5个不同成熟度。试验结果表明，本文模型较好地解决了自然复杂场景下番茄成熟度分级检测的错漏检问题，在模型浮点运算量（FLOPs）、参数量（Params）和内存占用量有少量增加的条件下，本文模型的平均精度均值为94.10%，相较原模型YOLO v8提高3.0个百分点。通过与Faster R-CNN-Resnet50、YOLO v5、YOLO v7-tiny、YOLO v8、YOLO v10和YOLO 11目标检测模型对比，本文在检测精度具有显著优势，为番茄成熟度的检测提供了一种可靠的方法。

Abstract:

The ripeness of tomatoes is closely related to their quality, and it serves as a crucial basis for key production processes such as harvesting and sorting. To address the issues of simple functionality in crop ripeness grading and detection systems, and high costs associated with manual system upgrades, taking tomatoes as an example. It collected and constructed a tomato image dataset under natural scenarios, and a semi-automatic tomato image annotation algorithm was designed based on the tomato fruit ripeness grading algorithm to annotate the collected data. Building on the YOLO v8 model, the FPN structure was replaced with the BiFPN structure to achieve more efficient multi-scale feature fusion. It utilized the SE attention mechanism for fused feature extraction across spatial and channel dimensions, and introduced the Focal SIoU loss function to measure the angular difference between the predicted bounding box and the ground truth box. This results in the development of the tomato ripeness grading and detection model YOLO v8_BFS which was based on color feature quantization and the improved YOLO v8, and can identify five different ripeness stages during tomato growth. Experimental results showed that the proposed model effectively solved the problems of false detection and missed detection in tomato ripeness grading and detection under complex natural scenarios. While there was a slight increase in model computational complexity (FLOPs), parameter count (Params), and memory storage size, the detection accuracy of the proposed model reached 94.10%, which was 3.0 percentage points higher than that of the original YOLO v8 model. Compared with target detection models such as Faster R-CNN-Resnet50, YOLO v5, YOLO v7-tiny, YOLO v8, YOLO v10, and YOLO 11, the proposed model demonstrated significant advantages in detection accuracy, providing a reliable method for tomato ripeness detection.