Application of Transfer Learning Techniques for Leaf Disease Detection of Images Under Controlled and Uncontrolled Environments

Abstract

Deep neural networks are pre-trained models based on the principle of transfer learning that have emerged as powerful tools for accurate and efficient disease detection in plant leaves. Pre-trained models are neural networks previously trained on vast and diverse datasets, often for image recognition tasks. Leveraging the knowledge learned from these pre-training tasks, they can be fine-tuned on smaller, domain-specific datasets for specific functions like plant leaf disease detection.

This paper aims to study the application of transfer learning models in crop disease classification with different experimental setups encompassing images under controlled and uncontrolled environments and compare the results. Images under controlled environments are taken in the lab setup with stable and precise background conditions. Images under uncontrolled environments are those captured with background noise,  varying light conditions, varying backgrounds, occlusion, etc. It is observed that irrespective of the model's sophistication, the performance of deep learning algorithms with images taken in a controlled environment outperforms those taken in uncontrolled environments. Hence, it is established that images captured in an uncontrolled environment should undergo necessary preprocessing to witness better performance of the deep learning algorithms.