基于YOLO v8-ABSeg的双孢蘑菇表型参数提取方法

Extraction Method of Phenotypic Parameters of Agaricus bisporus Based on YOLO v8-ABSeg

摘要:

针对双孢蘑菇采摘前人工获取其表型参数效率低、成本高的问题，提出了一种基于实例分割且适用于现代化工厂环境的双孢蘑菇表型参数提取方法。首先，对YOLO v8n-Seg实例分割模型进行改进，引入快速神经网络(Faster neural network，FasterNet)，并采用局部卷积(Partial convolutions，PConv)减少冗余计算和内存访问，引入SE（Squeeze-and-excitation）注意力机制到特征融合网络中，增加了网络对输入信息中重要部分的关注度，降低无关信息的干扰，改进后的模型完成了对双孢蘑菇目标的实例分割。最后，基于分割结果，提出了双孢蘑菇子实体4种表型参数的提取方法，包括菇盖直径、菇盖圆度、菇盖白度以及菇盖表面色斑。实验结果表明，YOLO v8-ABSeg模型在自建双孢蘑菇数据集上的mask精度比原模型提高了1.6个百分点，且参数量、浮点数运算量和内存占用量分别降低了38.7%、25.0%和36.8%，帧率提高了11.3%。此外，双孢蘑菇表型参数计算结果与人工测量结果误差小于10%。该方法可应用于双孢蘑菇表型参数的自动化获取，为生长模型建立、在线实时环境控制等提供技术基础。

Abstract:

In order to overcome the problem of low efficiency and high cost in the manual acquisition of phenotypic parameters of Agaricus bisporus, an instance segmentation-based method for calculating phenotypic parameters for modern industrial environments was proposed. Firstly, the YOLO v8n-Seg instance segmentation model was improved through the introduction of faster neural network (FasterNet), including the employment of partial convolutions (PConv) to reduce redundant computations and memory accesses. The squeeze-and-excitation (SE) attention mechanism was incorporated into the feature fusion network to enhance the model’s focus on the critical target components, minimizing interference from irrelevant background. The improved model successfully performed instance segmentation on Agaricus bisporus. Based on the segmentation results, four phenotypic parameters of the mushroom sub-entities were figured out: cap diameter, cap roundness, cap whiteness, and the color spots on the surface. Experimental results demonstrated that the YOLO v8-ABSeg model achieved a 1.6 percentage points improvement in mask accuracy on proposed custom-built Agaricus bisporus dataset, with reductions of 38.7%, 25.0%, and 36.8% in the number of parameters, floating-point operations, and weight file size, respectively，frames per second was increased by 11.3%. Additionally, the calculated phenotypic parameters exhibited a measurement error of no more than 10% when compared with manual measurement results. This method provided a foundation for the automation of phenotypic parameter extraction and can be applied to other applications like the development of growth models and real-time environmental control systems, and so on.