Electronic Theses and Dissertation

Klasifikasi tanaman herbal indonesia berbasis citra membutuhkan model yang tidak hanya akurat, tetapi juga memiliki kompleksitas rendah agar dapat diterapkan pada perangkat mobile. model deep learning berkapasitas besar umumnya mampu menghasilkan performa tinggi, tetapi memiliki jumlah parameter, ukuran penyimpanan, dan kebutuhan komputasi yang besar. penelitian ini bertujuan menerapkan knowledge distillation untuk meningkatkan performa model ringan pada klasifikasi tanaman herbal indonesia menggunakan dataset indoherb yang terdiri atas 100 kelas dan 10.000 citra. model teacher yang digunakan meliputi vit-base, swin transformer-base, beit-base, dan deit-base, sedangkan model student meliputi tinyvit-5m, deit-tiny, dan mobilevit-s. dua pendekatan distilasi diterapkan, yaitu response-based distillation dan feature-based distillation. evaluasi dilakukan menggunakan accuracy, macro precision, macro recall, dan macro f1-score, serta analisis efisiensi berdasarkan jumlah parameter, ukuran model, flops, latensi inferensi, throughput, dan penggunaan memori gpu. hasil penelitian menunjukkan bahwa swin transformer-base menjadi model teacher terbaik dengan accuracy 98,80% dan f1-score 98,79%. pada pelatihan tanpa distilasi, tinyvit-5m menjadi model student terbaik dengan accuracy 96,40% dan f1-score 96,37%. pendekatan response-based distillation menghasilkan performa terbaik pada kombinasi deit-base dan tinyvit-5m dengan accuracy 97,50% dan f1-score 97,52%. performa terbaik secara keseluruhan diperoleh melalui feature-based distillation antara swin transformer-base dan tinyvit-5m, dengan accuracy 98,10% dan f1-score 98,11%. model tinyvit-5m tersebut memiliki 5,10 juta parameter, ukuran 19,61 mb, dan kompleksitas 1,17 gflops, atau 94,12% lebih sedikit parameter dibandingkan swin transformer-base, dengan penurunan accuracy hanya 0,70 poin persentase. ketiga model student hasil distilasi berhasil dikonversi ke format onnx dan diimplementasikan pada prototipe aplikasi android. evaluasi tambahan menggunakan 260 citra dunia nyata dari 26 kelas menunjukkan bahwa tinyvit-5m hasil feature-based distillation memperoleh accuracy 95,77% dan f1-score 96,38%. hasil penelitian menunjukkan bahwa knowledge distillation efektif meningkatkan performa model ringan sekaligus mempertahankan efisiensi komputasi dan potensi implementasi praktis pada perangkat mobile.

Image-based classification of Indonesian herbal plants requires models that are not only accurate but also computationally efficient for deployment on mobile devices. Large-scale deep learning models generally achieve high classification performance but require substantial numbers of parameters, storage capacity, and computational resources. This study aims to apply Knowledge Distillation to improve the performance of lightweight models for Indonesian herbal plant classification using the IndoHerb dataset, which consists of 100 classes and 10,000 images. The teacher models include ViT-Base, Swin Transformer-Base, BEiT-Base, and DeiT-Base, while the student models include TinyViT-5M, DeiT-Tiny, and MobileViT-S. Two distillation approaches are evaluated: response-based distillation and feature-based distillation. Model performance is assessed using accuracy, macro precision, macro recall, and macro F1-score, while model efficiency is analysed based on the number of parameters, model size, FLOPs, inference latency, throughput, and peak GPU memory usage. The results show that Swin Transformer-Base is the best-performing teacher model, achieving 98.80% accuracy and a 98.79% F1-score. Without distillation, TinyViT-5M is the best student model, achieving 96.40% accuracy and a 96.37% F1-score. The best response-based distillation configuration is obtained using DeiT-Base as the teacher and TinyViT-5M as the student, achieving 97.50% accuracy and a 97.52% F1-score. The best overall performance is achieved through feature-based distillation between Swin Transformer-Base and TinyViT-5M, with 98.10% accuracy and a 98.11% F1-score. The distilled TinyViT-5M model contains 5.10 million parameters, has a model size of 19.61 MB, and requires 1.17 GFLOPs, representing 94.12% fewer parameters than Swin Transformer-Base while reducing accuracy by only 0.70 percentage points. All three distilled student models were successfully converted to the ONNX format and implemented in an Android-based application prototype. Additional evaluation using 260 real-world images from 26 classes shows that TinyViT-5M with feature-based distillation achieves 95.77% accuracy and a 96.38% F1-score. These findings demonstrate that Knowledge Distillation effectively improves the performance of lightweight models while maintaining computational efficiency and practical potential for mobile deployment.