A Lightweight RepViT-M1 Framework for Tea Leaf Disease Classification: Benchmarking Against VGG16 for Efficient Edge Deployment

Abstract

Exact and timely identification of tea (Camellia sinensis) leaf diseases is necessary for controlling yield and quality. However, field deployment is limited due to the large computational requirement of the standard deep convolutional networks. This paper presents a lightweight, deployment-oriented framework for classifying eight tea-leaf conditions from an imbalanced 885-image dataset and benchmarks it against a widely used VGG16 transfer-learning baseline. The baseline replaces the VGG16 classifier head and fine-tunes it on a class-balanced set, reaching 95.63% validation accuracy with a macro-averaged F1 of 0.957 but requiring roughly 134 million parameters. The proposed pipeline builds on a RepViT-M1 backbone (a mobile-oriented, parameterizable vision transformer) with a compact 128-dimensional bottleneck head, focal loss with label smoothing, differential learning rates, cosine annealing, and test-time augmentation. On a held-out test partition it attains 99.92% accuracy, a macro F1 of 0.999 and a one-vs-rest macro-ROC-AUC of 1.000 while using only about 4.8 million parameters, a reduction of roughly 28x relative to VGG16. An ablation over fine-tuning depth and a two-stage healthy-versus-diseased cascade are described as design options, and the trained model is exported to the ONNX interchange format (18 MB) to support cross-runtime inference on edge hardware. We report the results transparently, including a candid discussion of the class-balancing and evaluation protocol so that the efficiency gains are interpreted responsibly. The findings indicate that compact reparameterizable backbones can match far larger networks on tea-disease recognition, offering a practical route to on-device agricultural diagnostics.