基于多激发与层级标签的野生菌图像分类  

Classification of wild mushrooms images based on multi-excitation and hierarchical labels

作　　者：龙朝勋 李俊仪 李向阳 李海燕[1] 李红松[1] 余鹏飞[1] LONG Chao-xun;LI Jun-yi;LI Xiang-yang;LI Hai-yan;LI Hong-song;YU Peng-fei(School of Information Science and Engineering,Yunnan University,Kunming 650500,China)

机构地区：[1]云南大学信息学院,云南昆明650500

出　　处：《计算机工程与设计》2025年第4期941-949,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62066046);云南省重大科技专项计划基金项目(202202AD080004)。

摘　　要：为利用生物学多层级类别正确鉴别野生菌种类,提出一种多层级标签的分类网络。引入高效通道注意力(ECA)构建多激发模块,提取并融合多种判别性特征。推广条件概率权重矩阵并以决策融合构建多层级分类器(MHC),实现更广泛的逐层级监督与引导。基于准确率动态加权各层级损失项,调整训练侧重点,舍弃KL散度正则项,规避对交叉熵损失的干扰。实验结果表明,提出方法在“种”层级上Top1准确率可达98.17%,识别为可食用的有毒样本的比例(风险指数)达到最低水平,为3.64‰。t-SNE可视化显示,提取的特征的类内聚集性和类间可分离性有所提升。To correctly identify wild mushroom species using biological multi-hierarchy categories,a classification network with multi-hierarchy labels was proposed.Efficient channel attention(ECA)was introduced to construct multi-excitation modules to extract and fuse multiple discriminative features.The conditional probability weight matrix of the multi-hierarchy classifier(MHC)constructed by decision fusion was generalized to achieve a wider range of supervision and guidance.The classification accuracy of each level was used to dynamically weight corresponding loss term to adjust the focus of training,and the regularization term defined by KL divergence was discarded to avoid interference with cross-entropy loss.Results of experiments show that the proposed method achieves 98.17%in regard of Top1 accuracy at the species hierarchy.The proportion of poisonous samples in identified edible ones(Risk Index)reaches the lowest level at 3.64‰.The visualization with t-SNE shows that intra-class aggregation and inter-class separability of extracted features are well improved.

关 键 词：野生菌 层级标签 细粒度分类 决策融合 动态加权 KL散度 风险指数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]