一种精简的蘑菇图像分类模型  

A Light Mushroom Image Classification Model

作　　者：黄诗瑀 叶锋[1,2,3] 黄丽清 黄添强 陈家祯[1,2,3] 郑子华 HUANG Shiyu;YE Feng;HUANG Liqing;HUANG Tianqiang;CHEN Jiazhen;ZHENG Zihua(College of Computer and Cyber Security,Fujian Normal University,Fuzhou 350117,China;Digital Fujian Institute of Big Data Security Techonology,Fuzhou 350117,China;Fujian Provincial Engineering Research Center of Big Data Analysis and Application,Fuzhou 350117,China)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福建福州350117 [2]数字福建大数据安全技术研究所,福建福州350117 [3]福建省公共服务大数据挖掘与应用工程技术研究中心,福建福州350117

出　　处：《福建师范大学学报（自然科学版）》2023年第1期75-85,共11页Journal of Fujian Normal University：Natural Science Edition

基　　金：国家自然科学基金资助项目(62072106);福建省自然科学基金资助项目(2020J01168);福建省高校产学合作项目(2021H6004)。

摘　　要：相比小型卷积神经网络(convolutional neural network, CNN)模型,现有的大型CNN模型在大型图像数据集上达到了良好的分类效果,但是在小型图像数据集上过拟合,使得精度提升小、训练时间长、存储占用高,不能很好地适应嵌入式设备.因此首先收集了一个包含4 500张图片的小型蘑菇数据集,并为蘑菇分类任务设计了轻量化的CNN模型MushroomNet.然后研究CNN模型中各部分对于分类任务的重要性,并提出基于数据复杂度的模型结构优化方法.实验表明,相比MobileNet、ShuffleNet等轻量化模型,MushroomNet-MicroV2的Top-1精度只差了1%~2%,但是它训练速度更快,存储更小,只有1.3 M的参数量,且在Apple M1 CPU上经过142 s的30轮快速训练后,Top-1验证精度可达88%.Compared with small-scale convolutional neural network(CNN) models, existing large-scale CNN models have achieved good classification results on large image datasets.However, over-fitting on small image datasets results in little accuracy improvement, long training time and high storage consumption, so that they fail to adapt to embedded devices well.Therefore, this paper collects a small mushroom dataset with 4 500 images and designs a lightweight CNN model MushroomNet for the mushroom classification task.At the same time, this paper studies the importance of each part of the CNN model to the classification task and proposes a model structure optimization method based on data complexity.Experiments show that compared with lightweight models such as MobileNet and ShuffleNet, the Top-1 accuracy of MushroomNet-MicroV2 is only 1%~2% worse, but it has faster training speed and smaller storage, with only 1.3 M parameters.After 30 rounds of fast training in 142 s on the Apple M1 CPU, the Top-1 validation accuracy can reach 88%.

关 键 词：深度学习 模型裁剪 图像分类 卷积神经网络 计算机视觉 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]