基于PP-liteseg和Jetson Orin Nano的香菇菌棒表型测量方法

Phenotypic Measurement of Shiitake Mushroom Sticks Based on PP-liteseg and Jetson Orin Nano

摘要:

针对香菇育种过程中表型性状考察困难、人工测量费时费力的问题，提出一种适用于Jetson Orin Nano平台部署的香菇菌棒表型测量方法。使用手机和工业相机对3种不同数据来源的香菇菌棒进行图像采集，标注香菇菌落数据集并进行增强；比较了Bisenet、Stdcseg、U-net、Deeplabv3p、PP-liteseg 5个分割模型在Test-A、Test-B和Test-C测试集上的分割效果，结果表明PP-liteseg的普适性优于其他网络，PP-liteseg分割模型在3个测试集上的平均交并比超过97.53%，平均像素准确率高于99.49%，推理单幅图像耗时660 ms；为进一步平衡模型准确性和实时性，采用量化蒸馏方式对PP-liteseg模型进行压缩，并部署到Jetson Orin Nano平台上，压缩后的模型在Test-B测试集上的平均交并比和平均像素准确率分别为97.50%和99.51%，推理单幅图像耗时43.63 ms，比压缩前缩短近64%；采用PP-liteseg分割香菇菌棒图像，提取菌落表型，再根据菌落径向长度和轴向宽度得到菌丝生长长度，与人工测量值相比，菌丝生长长度平均绝对百分比误差、均方根误差和决定系数分别为1.874%、0.148 cm、0.918；采用该方法测量了6个菌株在连续4 d内的菌丝生长长度，结果表明，6个菌株的菌丝生长速度无论在单日还是在整个培养期内均有差异，其中49号和168号菌株差异最大。研究表明，本文方法适用于香菇菌棒表型测量，并能够以良好的准确性和实时性在Jetson Orin Nano平台上运行。

Abstract:

Aiming at the problems of difficulty in examining phenotypic traits and time-consuming and laborious manual measurements during shiitake mushroom breeding, a phenotypic measurement method for shiitake mushroom sticks applicable to the deployment of the Jetson Orin Nano platform was proposed. Image acquisition of shiitake mushroom sticks from three different data sources was performed by using cell phones and industrial cameras, and the shiitake mushroom colony dataset was labeled and enhanced;the segmentation effects of five segmentation models, namely, Bisenet, Stdcseg, U-net, Deeplabv3p, and PP-liteseg, were compared on the Test-A, Test-B, and Test-C test sets, and the results showed that PP-liteseg was the most suitable segmentation model for the test set of shiitake mushroom. The results showed that PP-liteseg was more pervasive than other networks, the mIoU of PP-liteseg segmentation models on the three test sets was more than 97.53%, the mPA was higher than 99.49%, and the inference of a single image took 660 ms. To further balance the accuracy and real-time performance of the model, the PP-liteseg model was compressed by quantized distillation and deployed on the Jetson Orin Nano platform, and the mIoU and mPA of the compressed model on the Test-B test set were 97.50% and 99.51%, respectively, and the inference of a single image took 43.63 ms, which was nearly 64% shorter than that of the pre-compression model;the images of shiitake mushroom sticks were segmented by using the PP-liteseg, and the phenotypes of colonies were extracted, and then the mycelial growth length was obtained based on the radial length and axial width of the colony, and the average absolute percentage error, root mean square error and coefficient of determination of the mycelial growth length were 1.874%, 0.148 cm and 0.918, respectively, compared with the manually measured values;the mycelial growth length of six strains were measured by the method in four consecutive days, and the results showed that the mycelial growth rate of six strains, both in a single day or during the whole culture period, with the greatest difference between strains 49 and 168. The study demonstrated that the method was suitable for the phenotypic measurement of shiitake mushroom mycelium and can be run on the Jetson Orin Nano platform with good accuracy and real-time performance.