Optimasi Arsitektur Jaringan Ringan untuk Pengenalan Palm Vein pada Perangkat Edge Menggunakan Hardware-Aware P-DARTS dan Knowledge Distillation

Fahmi, Mohammad Taris Syahir Zul (2026) Optimasi Arsitektur Jaringan Ringan untuk Pengenalan Palm Vein pada Perangkat Edge Menggunakan Hardware-Aware P-DARTS dan Knowledge Distillation. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025242008-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Palm vein recognition merupakan metode biometrik berbasis pola vena internal telapak tangan yang umumnya diperoleh melalui citra near-infrared (NIR). Untuk diterapkan pada perangkat edge seperti Raspberry Pi, sistem ini membutuhkan model yang akurat sekaligus efisien. Convolutional Neural Network (CNN) berkapasitas besar dapat memberikan representasi yang kuat, tetapi ukuran dan latency inferensinya menjadi kendala pada perangkat terbatas. Arsitektur CNN ringan telah menjadi pembanding kuat untuk efisiensi model. Namun, model pembanding tersebut tidak dirancang melalui pencarian yang secara eksplisit mempertimbangkan karakteristik dataset palm vein dan profil latency Raspberry Pi. Oleh karena itu, penelitian ini menerapkan hardware-aware Progressive Differentiable Architecture Search (P-DARTS) dengan penalti latency berbasis latency lookup table (LUT) Raspberry Pi, kemudian mengombinasikannya dengan Knowledge Distillation (KD) dan Post-Training Quantization (PTQ) INT8. Eksperimen menggunakan 8.340 citra dari SCUT_PV_v1 yang mencakup 834 kelas. Setelah tahap preprocessing, arsitektur dicari menggunakan hardware-aware P-DARTS dengan LUT Raspberry Pi. Genotype hasil pencarian kemudian menjalani retraining dan refinement untuk memperoleh kandidat student. EfficientNetV2M digunakan sebagai teacher pada KD. Model hasil KD diekspor ke ONNX dan dikonversi menjadi INT8 menggunakan PTQ untuk evaluasi pada Raspberry Pi 5. Hasil eksperimen menunjukkan bahwa kandidat hasil hardware-aware P-DARTS dengan koefisien penalti λ sebesar 0,05, lebar channel awal 12, dan 10 cell mencapai akurasi FP32 sebesar 98,92% setelah refinement. KD meningkatkan akurasi tersebut menjadi 99,76%. Setelah PTQ INT8, model final mencapai akurasi 99,64%, ukuran 0,928 MB, mean latency 3,87 ms, median latency 3,76 ms, dan latency p95 4,52 ms pada Raspberry Pi 5. Hasil ini menunjukkan bahwa integrasi hardware-aware P-DARTS, KD, dan PTQ INT8 memberikan kompromi yang seimbang antara akurasi, ukuran model, dan latency pada perangkat target.
===================================================================================================================================
Palm vein recognition is a biometric method based on internal vein patterns in the palm, commonly captured using near-infrared (NIR) imaging. Its deployment on edge devices such as Raspberry Pi requires an accurate and efficient model. High-capacity Convolutional Neural Networks (CNNs) can provide strong feature representations, but their model size and inference latency become constraints on resource-limited devices. Lightweight CNN architectures have become strong baselines for model efficiency. However, these baseline models were not designed through an architecture search process that explicitly considers the characteristics of the palm vein dataset and the Raspberry Pi latency profile. Therefore, this study implements hardware-aware Progressive Differentiable Architecture Search (P-DARTS) with a latency penalty based on a Raspberry Pi latency lookup table (LUT), combined with Knowledge Distillation (KD) and INT8 Post-Training Quantization (PTQ). The experiments used 8,340 images from SCUT_PV_v1, covering 834 classes. After preprocessing, architecture search was conducted using hardware-aware P-DARTS with the Raspberry Pi LUT. The resulting genotypes subsequently underwent retraining and refinement to obtain student candidates. EfficientNetV2M was used as the teacher during KD. The distilled model was exported to ONNX and converted to INT8 through PTQ for evaluation on Raspberry Pi 5. The hardware-aware P-DARTS candidate obtained with a latency penalty coefficient λ of 0.05, an initial channel width of 12, and 10 cells achieved an FP32 accuracy of 98.92% after refinement. KD increased the accuracy to 99.76%. After INT8 PTQ, the final model achieved 99.64% accuracy, a model size of 0.928 MB, a mean latency of 3.87 ms, a median latency of 3.76 ms, and a p95 latency of 4.52 ms on Raspberry Pi 5. These results demonstrate that integrating hardware-aware P-DARTS, KD, and INT8 PTQ provides a balanced trade-off among accuracy, model size, and latency on the target device

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Hardware-Aware P-DARTS, Latency Lookup Table, Knowledge Distillation, Palm Vein Recognition, Post-Training Quantization,
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Mohammad Taris Syahir Zul Fahmi
Date Deposited:	31 Jul 2026 02:19
Last Modified:	31 Jul 2026 02:19
URI:	http://repository.its.ac.id/id/eprint/139434

Actions (login required)

View Item