Article Per Year (5 Year)
p-Index From 2021 - 2026
0.23
P-Index
This Author published in this journals
All Journal SKANIKA: Sistem Komputer dan Teknik Informatika
Nur Dwi Priyambodo
Universitas Dian Nuswantoro
Computer Science & IT Control & Systems Engineering Engineering

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

 1 
OPTIMASI MOBILENETV2 DENGAN PRUNING DAN QUANTIZATION UNTUK DETEKSI PENYAKIT DAUN PADI PADA PERANGKAT EDGE Nur Dwi Priyambodo; Edi Sugiarto
SKANIKA: Sistem Komputer dan Teknik Informatika Vol 9 No 2 (2026): Jurnal SKANIKA Juli 2026 (In Press)
Publisher : Universitas Budi Luhur

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.36080/skanika.v9i2.3828

Abstract

Rice leaf diseases Bacterial Leaf Blight (BLB), Brown Spot, and Leaf Blast pose a serious threat to food security, with potential yield losses of 30–70%. Existing deep learning models are typically large (>80 MB), making them impractical for farmers' low-specification smartphones. This study proposes MobileNetV2 optimization through 40% magnitude-based weight pruning and Post-Training Quantization (PTQ) Int8 using TensorFlow Lite for Android deployment. The dataset is sourced from a Mendeley repository (DOI: 10.17632/fwcj7stb8r/1) comprising 5,932 images across 4 classes, from which a 3-class subset (BLB, Brown Spot, Leaf Blast) totaling 4,804 images was used, split 70-15-15 into training, validation, and test sets. Evaluation was performed on 721 held-out test images, excluded from the training, validation, and quantization calibration processes. The proposed model achieves 2.89 MB (88.23% reduction from the 24.58 MB baseline), an inference speed of 14.53 ms/image on an Intel CPU workstation, and an accuracy of 98.34% (macro F1-Score 0.9831). More aggressive 50% pruning caused accuracy to degrade to 83.50% post-quantization, confirming 40% as the best trade-off among the sparsity levels tested. Grad-CAM analysis validated the biological relevance of the extracted features for each disease class. Future work includes on-device validation on Android hardware and expanded disease class coverage
Co-Authors Edi Sugiarto