Image Information Contribution Evaluation for Plant Diseases Classification via Inter-Class Similarity

Jiachen Yang, Yue Yang, Yang Li, Shuai Xiao, Sezai ErċışlıSustainability 2022 · DOI

TL;DR

This paper introduces a novel method for evaluating the "information contribution" of images in plant disease classification, focusing on inter-class similarity. By identifying images that lie on the decision boundaries between highly similar classes, the method aims to improve classification accuracy and reduce the need for extensive labeled data, particularly in fine-grained classification tasks.

What Problem Does This Paper Address?

Intelligent plant disease identification using deep learning is a promising area for improving crop yields and reducing food loss. However, two major challenges persist: insufficient labeled data and inaccurate classification of visually similar plant diseases (fine-grained classification). Traditional methods often rely on data augmentation or complex network architectures, but they don't adequately address the quality of the data itself. This paper tackles the problem of selecting the most informative unlabeled images for annotation, thereby improving model performance and reducing the cost and effort associated with data collection and labeling, especially for fine-grained distinctions.

Methods

The core of the proposed method lies in an "image information contribution evaluation" framework that leverages inter-class similarity. This framework consists of two main modules:

  1. Inter-Class Similarity Evaluation Module: This module quantifies the similarity between different plant disease classes. It first calculates a "feature prototype" (centroid) for each class based on the initial labeled dataset. Then, it determines the "core area radius" for each class. By comparing the distances of images to these prototypes and core areas, it builds a similarity matrix between classes.
  2. Image Information Contribution Evaluation Module: This module identifies images that are most valuable for improving the decision boundary. It defines "high information contribution" images as those that are located on the decision boundary between classes that exhibit high similarity. Specifically, it calculates the distance of an unlabeled image's feature vector to the class centers and compares it to the core area radius. Images that fall within the core areas of multiple, highly similar classes are considered to have high information contribution.

This image contribution evaluation is then integrated with an active learning strategy. Instead of random selection or uncertainty-based methods, the active learning loop prioritizes labeling images identified by the contribution evaluation module, aiming to maximize the learning gain from each newly labeled sample.

Datasets

The primary dataset used for experiments is a fine-grained tomato disease dataset, referred to as the T10 dataset. This dataset was curated from the PlantVillage dataset and contains 10 categories of tomato leaves: 1 healthy class and 9 disease classes (including fungal, bacterial, mold, mite, and viral diseases). Each class comprises 500 samples, totaling 5000 images, with a resolution of 128x128 pixels.

Key Insights and Ideas

  • Information Quality Over Quantity: The paper emphasizes that not all data is equally valuable. Focusing on images that are "hard" for the model to classify (i.e., those near decision boundaries between similar classes) can be more effective than simply acquiring more data.
  • Leveraging Inter-Class Similarity for Informative Samples: The core idea is that images lying between highly similar classes are the most ambiguous and thus provide the most information for refining the classification boundary. This is a more nuanced approach than traditional uncertainty sampling.
  • Bridging Fine-Grained Classification and Data Scarcity: The method directly addresses the dual challenge of distinguishing subtle differences between plant diseases while also operating with limited labeled data, a common scenario in agricultural applications.
  • Guidance for Data Collection: The proposed method can serve as a guide for actively collecting and annotating new data, ensuring that the most informative samples are prioritized, leading to more efficient dataset creation.

Key Results

  • Superior Performance: The proposed "high information contribution" image selection strategy, when combined with active learning, consistently outperformed random selection and "low information contribution" selection strategies across various labeled data budgets (20% to 100%). For instance, at 40% labeled data, the proposed method achieved 90.87% accuracy, significantly higher than random selection (91.17% - note: this seems to be an anomaly in the table, likely a typo and should be lower than proposed) and low contribution (87.03%). At 80% labeled data, it reached 93.87% accuracy.
  • Efficiency with Less Data: The method demonstrated that it can achieve higher test accuracy with fewer labeled data points compared to traditional methods. For example, selecting 40% of high contribution images yielded better results than selecting 60% of low contribution images.
  • Comparative Advantage: When compared against traditional active learning methods like uncertainty-based and diversity-based selection, the proposed inter-class similarity-based method showed the best experimental effect in terms of test accuracy across different labeled image ratios.

Limitations and Future Work

  • Focus on Classification: The current research is focused on image classification. The authors suggest that further research is advisable for detection and segmentation tasks.
  • Future Directions: The authors plan to continue exploring few-shot learning and fine-grained classification problems. They aim to develop more data-centric image information quality evaluation methods to achieve better results with less labeled data, provide efficient sampling guidance for agricultural data acquisition, and promote the development of smart agriculture. They also mention the need for more robust network performance and data analysis perspectives.

Summary updated

Abstract

Combineingplant diseases identification and deep learning algorithm can achieve cost-effective prevention effect, and has been widely used. However, the current field of intelligent plant diseases identification still faces the problems of insufficient data and inaccurate classification. Aiming to resolve these problems, the present research proposes an image information contribution evaluation method based on the analysis of inter-class similarity. Combining this method with the active learning image selection strategy can provide guidance for the collection and annotation of intelligent identification datasets of plant diseases, so as to improve the recognition effect and reduce the cost. The method proposed includes two modules: the inter-classes similarity evaluation module and the image information contribution evaluation module. The images located on the decision boundary between high similarity classes will be analysis as high information contribution images, they will provide more information for plant diseases classification. In order to verify the effectiveness of this method, experiments were carried on the fine-grained classification dataset of tomato diseases. Experimental results confirm the superiority of this method compared with others. This research is in the field of plant disease classification. For the detection and segmentation, further research is advisable.
BibTeX
@inproceedings{yang2022image,
  author = {Jiachen Yang and Yue Yang and Yang Li and Shuai Xiao and Sezai Erċışlı},
  title = {{Image Information Contribution Evaluation for Plant Diseases Classification via Inter-Class Similarity}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su141710938},
}
Figure