基于轻量化YOLO v5s-MCA的番茄成熟度检测方法

Tomato Maturity Detection Method Based on Lightweight YOLO v5s-MCA

摘要:

针对自然环境下番茄识别易受复杂背景干扰、相邻果实成熟度相似难以检测等问题，本文提出了一种轻量化YOLO v5s-MCA番茄成熟度识别模型，划分成熟期、转熟期、转色期和未熟期4个成熟度等级。该模型在YOLO v5s基础上使用MobileNetV3网络，减少了模型参数量；在主干网络和颈部网络引入坐标注意力机制（Coordinate attention，CA），提高了模型对番茄特征表达能力；将颈部网络替换为加权双向特征金字塔网络BiFPN，强化了模型特征融合性能并提高了模型识别准确率；将颈部网络中的标准卷积模块改进为GSConv卷积，减轻了模型复杂度并提高了对目标信息的获取能力。试验结果表明，YOLO v5s-MCA模型参数量仅为2.33×106，计算量仅为4.1×109，模型内存占用量仅为4.83 MB，其精准度和平均精度均值分别达到92.8%和95.1%，相对YOLO v5s基础模型分别提升3.4、4.4个百分点。对比YOLO v3s、YOLO v5s、YOLO v5n、YOLO v7、YOLO v8n及YOLO v10n等6种模型，YOLO v5s-MCA模型轻量化效果与检测性能最优。

Abstract:

Aiming to address the challenges of tomato recognition in natural environments, such as interference from complex backgrounds and difficulty in detecting adjacent fruits with similar ripeness levels, a lightweight YOLO v5s-MCA model for tomato ripeness detection was proposed. The model categorized tomato ripeness into four distinct stages: mature, turning mature, color transition, and immature. Firstly, it incorporated the MobileNetV3 network as the backbone, significantly reducing the model’s parameter count and computational requirements. Moreover, the coordinate attention (CA) mechanism was integrated into the backbone and neck networks, enhancing the model’s ability to enhance the model’s ability to represent tomato features. Additionally, the neck network was replaced with a weighted bidirectional feature pyramid network (BiFPN) to strengthen feature fusion and improve recognition accuracy. The standard convolution modules in the neck network were also replaced with GSConv convolution to reduce model complexity and enhance the ability to capture target information. Experimental evaluations revealed the superior performance of the YOLO v5s-MCA model. The model achieved a parameter count of only 2.33×106, with a computational cost of 4.1×109 and a memory footprint of just 4.83 MB. The model achieved a precision of 92.8% and a mean average precision (mAP) of 95.1%, representing improvements of 3.4 percentage points and 4.4 percentage points, respectively, compared with the baseline YOLO v5s model. To further validate the effectiveness of the YOLO v5s-MCA model, it was compared with six other models, including YOLO v3s, YOLO v5s, YOLO v5n, YOLO v7, YOLO v8n, and YOLO v10n. Among these, the YOLO v5s-MCA model outperformed its counterparts in terms of lightweight design and detection performance.