Unsupervised deep metric learning algorithm for crop disease images based on knowledge distillation networks

文献类型： 外文期刊

第一作者： Zeng, Qingtian

作者： Zeng, Qingtian;Li, Xinheng;Wang, Shansong;Ni, Weijian;Duan, Hua;Xie, Nengfu;Xiao, Fengjin

作者机构：

关键词： Unsupervised deep metric learning; Crop disease images; Metric spaces; Distillation networks

期刊名称：MULTIMEDIA SYSTEMS （ 影响因子：3.1； 五年影响因子：3.3 ）

ISSN： 0942-4962

年卷期： 2024 年 30 卷 5 期

页码：

收录情况： SCI

摘要： Current crop disease image classification models usually require a large number of pre-labeled training images for training, however, obtaining crop disease image datasets with accurate labels is a challenging task. Many studies have addressed this problem based on unsupervised deep metric learning (UDML), which exploits the similarity of unlabeled crop disease images in the embedding space. However, the successful implementation of UDML methods relies on accurate pseudo labels to guide model training, which may prevent the model from converging correctly during the learning process if the pseudo labels has large errors. To overcome these limitations, we propose a UDML-based algorithm for classifying crop disease images called an unsupervised deep metric learning algorithm based on a knowledge distillation network (KDUM). The KDUM algorithm can extract accurate pseudo labels from unlabeled crop disease images to guide the training process of the model. Specifically, we map the crop disease images to the metric space through the mapper of the teacher network. A clustering algorithm is used in the metric space to generate hard pseudo labels with noise for computing the loss in the metric space. At the same time, the classifiers of the teacher network are used to generate robust soft pseudo labels as auxiliary training information. The information is used with the output of the classifiers of the student network to calculate the soft cross-entropy loss, which is used to assist in guiding the training of the student network. When updating the model parameters, an exponentially weighted moving average is used to smoothly update the parameters of the teacher network. Mitigating the problem of error amplification caused by noisy pseudo labels generated by clustering. We conducted comparative experiments with seven previous years' state-of-the-art (SOTA) methods on three datasets. The results show that our algorithm improves the NMI metrics by about 3.61% on the Plant Village dataset, 3.09% on the AFL-33 dataset, and 7.76% on the Kaggle Tomato dataset compared to the SOTA methods.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序