Eye-Controlled Plant Disease Diagnosis for Severe Motor Disabilities Using Transfer Learning

Abstract

Plant disease diagnosis systems commonly require manual hand interaction, which creates accessibility barriers for individuals with severe motor disabilities, including farmers and students who cannot use their hands. This paper presents a low-cost, hands-free system that enables users to control a multi-crop disease diagnosis interface using only eye gaze and blinking. The proposed system integrates three main components: real-time eye tracking using MediaPipe (achieving ninety-eight percent face detection and eighty-five percent blink detection accuracy), a MobileNetV2 deep learning model converted to TensorFlow Lite that achieves 98.22% validation accuracy across eight disease classes, and a visual similarity verification module using the Structural Similarity Index (SSIM). A user study was conducted with fifteen participants, including two individuals with motor disabilities. Results indicated that all fifteen participants agreed that the system would benefit individuals with motor disabilities. The similarity verification module provided an additional validation layer for classification results and confirmed the consistency of predictions for visually similar plant disease samples. The system operates in real-time (thirty frames per second) on a standard laptop with a webcam and provides an open-source framework for accessible agricultural technology.