AUTHOR=Lin Hong , Qiang Zhenping , Tse Rita , Tang Su-Kit , Pau Giovanni 

TITLE=A few-shot learning method for tobacco abnormality identification

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 15 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2024.1333236

DOI=10.3389/fpls.2024.1333236

ISSN=1664-462X

ABSTRACT=Tobacco is a valuable crop but its diseases identification is rarely involved in existing works.In this work, we use few-shot learning (FSL) to identify abnormalities of tobacco. FSL is a solution for the data deficiency that has been an obstacle of using deep learning. However, weak feature representation caused by limited data is still a challenging issue of FSL. The weak feature representation leads to weak generalization and troubles of cross-domain. In this work, we propose a feature representation enhancement network (FREN) which enhances the feature representation in instance-embedding and task-adaptation. For instance-embedding, global max-pooling and global avg-pooling are used together for adding more features, Gaussian-like calibration is used for normalizing the feature distribution. For task-adaptation, self-attention is adopted for task contextualization. Given the absence of publicly available data on tobacco, we create a tobacco leaf abnormality dataset (TLA), which includes 16 categories, 2 settings and 1430 images totally. In experiments, we use PlantVillage that is the benchmark dataset for plant disease identification to validate the superiority of FREN firstly. Then we use the proposed method and TLA to analyze and discuss the abnormality identification of tobacco. For the multi-symptom diseases that always have low accuracy, we propose a solution by dividing the samples into sub-categories created by symptom. For the 10 categories of tomato in PlantVillage, the accuracy achieves at 66.04% in 5-way, 1-shot tasks. For the two settings of tobacco leaf abnormality dataset, the accuracies achieve at 45.5% and 56.5%. By using multi-symptom solution, the best accuracy can be lifted to 60.7% in 16-way,1-shot tasks, and achieves at 81.8% in 16-way, 10-shot tasks. The results show that our method improves the performance greatly by enhancing feature representation, especially benefits for tasks that contain categories with high similarity.The desensitization of data when crossing domains also validates that the FREN has strong generalization ability.