<![CDATA[Diseases and pest infestations on tea plants can significantly reduce the quality and quantity of production, necessitating an early detection system based on artificial intelligence. Although deep learning architectures are capable of providing high accuracy, their large model size remains a major constraint for deployment on resource-limited devices. This study aims to compress the ShuffleNet V1 architecture using an L2-Norm based structured pruning method for the classification of six classes of tea leaf conditions utilizing a dataset from Mendeley Data. The model evaluation is carried out using the 5-Fold Cross Validation method with a fine-tuning process for 5 epochs to restore the network representation capacity after pruning. The experimental results demonstrate that the application of structured pruning successfully reduces the total parameters and computational operations significantly without sacrificing model performance. The compressed model is able to maintain an optimal accuracy reaching up to 99% across various pruning scale scenarios from 10% to 50%, while simultaneously providing a noticeable inference speedup. In the final stage, the best compressed model file is integrated into a web interface program, enabling users to perform tea leaf disease classification practically and responsively directly through a web browser. This research proves that the combination of ShuffleNet V1 and structured pruning can produce a highly lightweight yet accurate model for web implementation needs]]>
Copyrights © 2026
