基于GAN网络的菌菇表型数据生成研究

生成式对抗网络是基于对抗过程生成数据模型的新框架，它能够生成高质量的图像数据，为解决小样本数据、非均衡数据分析等提供了行之有效的方法。菌菇作为重要的真菌之一，其种类繁多，数据长尾分布、非均衡性等为其表型智能化识别与分类带来了困难。针对蘑菇表型数据，设计了一个高效的蘑菇表型生成式对抗网络MPGAN。研究了菌菇表型数据生成技术，设计了用于菌菇表型数据生成的生成式对抗网络结构，系统分为模型训练和表型图像生成两个模块。为了提升生成质量，使用Wasserstein距离和带有梯度惩罚的损失函数。利用开源数据和私有数据集测试学习率、处理所需的批次数EPOCH与Wasserstein距离。系统生成的菌菇表型数据为后期菌菇数据分类与识别提供了大数据基础。

Mushroom Phenotypic Generation Based on Generative Adversarial Network

Phenotypic data analysis based on image data and machine learning has become one of the important issues in interdisciplinary research. In recent years, the big data and deep learning techniques have provided powerful tools for image analysis and machine vision. Currently, the generative adversarial network is becoming a novel framework for the process estimation generation model. It can generate high quality image data and provide an effective approach for solving the problem of small sample data and unbalanced data analysis and so on. As one of the important fungi, mushroom has a plenty of varieties and the long tail distribution and non equilibrium of the data distribution bring great difficulties to its phenotypic intelligent classification and identification. Aiming to design a high efficiency mushroom phenotype resistance network MPGAN with mushroom phenotype data. The phenotypic data generation technology of mushroom was studied, and the generated confrontation network structure for mushroom phenotypic data generation was designed. The system was divided into two modules: model training and phenotypic image generation. To improve the quality of the generation, Wasserstein distances and loss functions with gradient penalty were used. Experiments were conducted on two datasets: open source data and private data sets, and results analysis were performed with the learning rate, number of batches required to process EPOCH and Wasserstein distances. The phenotypic data of the mushroom produced with this approach can furnish data basis for the classification of the mushroom data in the later stage, and provide solutions for solving the issues of unbalanced data and long tail distribution of the mushroom classification. The research can provide technical support for the study of high quality mushroom phenotypic data sets.