Implementation of real time root crop leaf classification using CNN on raspberry-Pi microprocessor

Abstract

This work presents the implementation of deep learning models for classifying root crop leaves, specifically beetroot, potato, radish, and sweet potato. Applying ResNet50 and DenseNet121 architectures, the work demonstrates the classification based on a comprehensive dataset of over 2,500 images collected from various locations across Karnataka, India. Both models exhibited good performance, with ResNet50 achieving 99.60 % accuracy and DenseNet121 attaining 97.60 %. The models maintained high precision, recall, and F1 scores across all classes, using CPU. A key achievement was the successful deployment of these models on a Raspberry Pi 4B, with ResNet50 maintaining its high accuracy with 99.60 % and DenseNet121 achieving 96.81 % accuracy on this resource constrained device. The practical applicability was further validated through field testing, where the Raspberry Pi 4B setup was mounted on a vehicle with the webcam to capture root crop leaves in real time and successfully tested in actual agricultural field. This demonstrated the system's viability for real-time crop classification. The outcomes highlight the promise of deep learning models in agriculture technology by providing a dependable, effective, and portable method for classifying root crop leaves. The consistent high accuracy maintained across different hardware platforms and in real-world conditions demonstrates the robustness and versatility of the developed models.