LeafLiteX mobile application for leaf disease detection using U-Net segmentation and lightweight deep learning.

Abstract

Agriculture is significant in world food production and global economic stability, but leaf disease and pest infection can cause a threat to crop quantity and quality. Thus, it became crucial to have timely and accurate identification of plant leaf disease to prevent loss in agriculture on a large scale and to have sustainable crop management. This paper introduces LeafLiteX, a lightweight mobile-based deep learning application for real-time detection and classification of leaf diseases. This application uses U-Net segmentation to precisely find leaf regions and MobileNetV3-Large to quickly classify diseases with less computation on the computer. The application performs end-to-end processing, from image acquisition to segmentation and disease prediction on mobile devices. An experiment was performed on publicly available crop disease datasets containing various leaf images having different disease types. The model obtained an accuracy of 98.85% showing improved generalization with minimal latency. The design of the model was such that it was suitable for inference on-device while still being robust enough despite changes in lighting conditions, background noise, and camera resolution. LeafLiteX is a low-cost, easy to use, offline-capable, and in-the-moment decision-making supportive diagnostic application that supports farmers and agrarians who require early detection. This paper demonstrates the capabilities that can be achieved using edge-optimized machine learning and computer vision to support the development of smart agriculture technologies. While traditional methods rely solely on classification, this research focuses more on practical implementation by incorporating segmentation, lightweight classification, and explainability to develop a mobile-friendly model.