International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 11 Issue: 12 | Dec 2024
p-ISSN: 2395-0072
www.irjet.net
Effective Plant Disease Identification Using Few-Shot Learning Sujay Maladi1, Hrishikesh Akundi2, Pallavi Vanamada3, Rahul Bellamkonda4 1234Department of Computer Science and Engineering, GITAM (Deemed to be University),
Visakhapatnam, Andhra Pradesh, India ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Plant diseases present a major problem to the
solutions for detecting and classifying plant diseases but most of these approaches are impractical due to the need of collecting and annotating large datasets of plant images. Hence, this study uses an approach called Few-Shot Learning.
agriculture sector, significantly impacting crop yields. Timely detection of plant diseases with good precision and accuracy is crucial for controlling outbreaks and reducing their effects on agricultural production. Although there have been recent developments in the use of Artificial Intelligence and Machine Learning for the detection of plant diseases, these solutions are dependent on large datasets of annotated images which may not be readily available in real life. This study examines the application of Few-Shot Learning approach for plant disease identification in scenarios with limited datasets. Few-Shot Learning enables a model to recognize new classes with very small number of examples, making it highly efficient. It imitates human learning abilities to generalize from minimal examples. This contrasts traditional machine learning methods, which usually require huge amounts of labeled data to achieve good accuracy. In this paper, the PlantVillage dataset has been used and EfficientNet, a pre-trained Convolutional Neural Network which was trained on the ImageNet dataset has been used along with Siamese Networks and a Triplet Loss function for FSL. The proposed approach is tested and compared with the regular fine-tuning transfer learning. The outcomes reveal that Few-Shot Learning with EfficientNet and Siamese Networks outshines the conventional transfer learning methods for the detection of plant diseases especially in the scenarios where quick adaptability is needed.
1.1 Few-Shot Learning Few-Shot Learning is a machine learning approach where a model learns to make predictions by training with only a very small number of examples. This approach is mainly used when data is limited and the model needs to adjust efficiently to new tasks. Few-Shot Learning is generally used in applications like face recognition, medical imaging, natural language processing, visual search, robotics, game AI, etc. Few-Shot Learning can be studied using the N-way-K-shot structure where N indicates the number of classes and K indicates the number of examples provided for each of the N classes. For example, if it is 4-way-2-shot task, it means it contains 4 classes of images and 2 examples of each class are provided. The higher the N value, the more difficult the task is but the higher the K value, the easier the task gets because more supporting information is available to derive an inference. In Few-shot learning, the K value is typically less than or equal to 10. Few-shot learning is given specific names when K = 0 and K = 1. If K = 0, it is called zero-shot learning, and if K = 1, it is called one-shot learning.
Key Words: Machine Learning, Deep Learning, Few-shot Learning, Transfer Learning, Fine-tuning, Siamese Networks, Triplet Loss, EfficientNet
2. RELATED WORK
1. INTRODUCTION
Deep learning approaches based on convolutional neural networks (CNNs) have proven highly effective for plant disease classification as highlighted in [3, 4, 5, 6]. However, with the rapid emergence of new plant diseases, there is a need for plant disease detection technologies that can quickly adjust to changes and operate effectively with limited data [7, 8]. In this context, few-shot learning appears to be a good solution.
The agriculture sector is the backbone in providing food for human beings and animals, as well as in helping achieve sustainable development goals, especially SDG 2: Zero Hunger. The agriculture sector mainly deals with plants, and currently, plant diseases are a major problem. Plant diseases that are not detected and managed early on may spread throughout the entire crop. This not only reduces produce quality but also causes huge economic losses for farmers and may harm the ecosystem. Manual plant disease detection is susceptible to human error and demands extensive knowledge in the field.
In [9] the authors proposed a few-shot learning approach using Siamese Networks. Here, the features are extracted by a two way convolutional neural network (CNN) with shared weights, a spatial structure optimizer (SSO) enhanced the metric learning and a k-nearest neighbor (kNN) classifier is used for classification. In [10] a low-shot learning method was proposed, where support vector machine (SVM) was employed to segment disease spots while preserving edge
So, there is a clear need for automated plant disease detection system [1]. Images of plant leaves can be used to easily detect the presence of a disease and identify its type [2]. Artificial Intelligence and Machine Learning have introduced many
© 2024, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 887