论文提出一种新的植物叶片分类以及分级网络模型MGMS(multi-granularity and multi-stage network)，该模型关注叶片的多粒度信息，并将多层级粒度特征进行有效融合。模型骨架由通用特征提取网络Resnet50构成，在不同阶段计算提取出特征，并将这些特征拼接，完成多粒度特征提取。此外，还使用了中心差分卷积模块，使模型可以关注图像中更具区分度的局部区域，提取出的特征更具区分性。在训练时采用多阶段训练方式，通过计算每一步提取的特征(包括拼接特征)得到的预测标签与真实标签的损失函数，实现由浅层特征到深层特征的学习，将triplet loss用于模型训练，通过减小an-chor与正样本的欧式距离，增大anchor与负样本欧式距离优化目标。该方法在Flavia leaf和Swedish leaf两个公开的叶片分类数据集上分别达到99。3%和99。9%的分类准确率，其中在Swedish leaf数据集上达到了目前最高准确率，在Flavia leaf上与当前最高准确率的方法相当，且在构建的烟叶分级数据集上也达到目前最高的71。2%的分级准确率。
A Leaf Classification and Ranking Method Based on Multi-granularity Multi-stage Feature Learning
This article proposes a new plant leaf classification and grading network model MGMS(Multi Granularity and Multi Stage network),which focuses on the multi granularity information of leaves and effectively integrates multi-level granularity fea-tures.Specifically,the model skeleton is composed of the general feature extraction network Resnet50,which calculates and ex-tracts features at different stages,and concatenates these features to complete multi-granularity feature extraction.In addition,a central differential convolution module is used to enable the model to focus on more discriminative local regions in the image and ex-tract more distinctive features.During training,a multi-stage training approach is adopted to achieve learning from shallow features to deep features by calculating the loss function between the predicted labels obtained from each extracted feature(including concat-enated features)and the true labels.The article uses triplet loss for model training and increases the optimization objective of the Eu-clidean distance between anchors and negative samples by reducing the Euclidean distance between anchors and positive samples.This method achieves classification accuracies of 99.3%and 99.9%on two publicly available leaf classification datasets,Flavia Leaf and Swedish Leaf respectively.The highest accuracy is achieved on the Swedish Leaf dataset,which is comparable to the cur-rent highest accuracy method on Flavia Leaf.Additionally,the highest classification accuracy of 71.2%is achieved on the construct-ed tobacco grading dataset.
leaf classificationmulti-granularity fusiontraining by levelcentral difference convolution
万方数据
维普
