A Scalable ONNX-Based Hybrid CNN Framework for Real-Time Crop Disease Detection Using EfficientNetB0

Abstract

Agricultural productivity is significantly affected by crop diseases that reduce yield quality and quantity. Early detection of plant diseases remains a major challenge due to variations in environmental conditions, visual similarity among diseases, and limited access to expert knowledge in rural regions. Manual inspection processes are often time-consuming, inconsistent, and prone to human error, leading to delayed intervention and increased crop damage. The proposed system introduces an intelligent crop disease detection framework based on a hybrid convolutional neural network utilizing EfficientNetB0 for accurate multi-class classification of plant diseases. The system processes leaf images from Pepper, Potato, and Tomato crops and classifies them into fifteen distinct categories, including healthy and diseased conditions. A structured pipeline involving data collection, preprocessing, model training, and evaluation ensures robustness and reliability. Integration with an optimized inference mechanism enables real-time prediction with reduced computational overhead. The framework provides outputs including crop type, disease category, confidence level, and severity estimation, supporting informed agricultural decision-making. Enhanced generalization is achieved through diverse dataset representation and augmentation strategies. Experimental results demonstrate high classification performance and efficient execution, making the system suitable for deployment in resource-constrained environments. The proposed solution contributes to smart agriculture by enabling rapid disease identification, reducing dependency on manual expertise, and improving crop management practices for sustainable farming.