University of the Philippines Diliman • Preprint 2026

Real-Time Baybayin Script Detection on Edge Devices

Preserving cultural heritage with post-training quantized INT8 YOLO26. Achieving sub-45ms latency and ultra-lightweight 1.8MB weights for web, mobile, and offline embedded execution.

Open Live Recognition Demo Kaggle Dataset & Weights
JN

James Arnold Nogra*

Data Science Program, UP Diliman

RS

Rachelle Sambayan

Institute of Mathematics, UP Diliman

Model Compactness
1.8 MB

Nano ¼ INT8 weights; 50% footprint reduction without losing architectural integrity.

KB
Edge Latency
45.1 ms

Tested on iPhone 16 Pro (Web/TensorFlow.js/WASM). Under 240ms on older devices.

MS
Detection Precision
0.7943

Nano ½ mAP50-95 with only a negligible 0.0064 delta drop from FP32 baseline.

mAP
Character Diversity
59 Classes

Covers native letters, kudlit variations, and virama diacritics in 3,094 annotations.

ᜃ

Empirical Evaluation

Precision vs. Footprint Benchmark

Results from 5-fold stratified cross-validation on multi-object composite capture boards comparing full FP32 vs. Post-Training INT8 Quantization.

Detection Accuracy (mAP50-95)

FP32 Baseline vs. Quantized INT8

Table III

Memory Footprint (KB)

Storage reduction after Symmetric INT8 clipping

Table II

Device Inference Latency (ms)

Averaged across 25 real-world inference runs (warm-run timings)

Up to 2x Faster with INT8
Device Target Nano 1/16 (INT8) Nano 1/4 (INT8) Nano 1/2 (INT8) Small (INT8) Small (FP32)
iPhone 16 Pro 45.10 ms 47.00 ms 54.00 ms 57.40 ms 124.60 ms
Samsung Fold 6 80.55 ms 87.64 ms 92.70 ms 110.28 ms 144.42 ms
Ryzen 7 8745H 57.72 ms 57.72 ms 59.66 ms 76.04 ms 196.46 ms
iPhone 11 (2019) 155.80 ms 161.60 ms 183.40 ms 238.80 ms 240.60 ms
Methodology

Pruned Nano Backbones & MuSGD Optimization

To enable low-power processing without specialized hardware, we introduced four custom channel configurations (Nano ½, Nano ¼, Nano ⅛, and Nano ¹/₁₆) downscaling feature extraction channels while retaining C3k2, SPPF, and C2PSA attention blocks.

1

MuSGD Optimizer: Trained using Stochastic Gradient Descent with Muon optimization combining momentum dynamics and Newton-Schulz matrix factor adjustments.

2

Symmetric Quantization: Weights scaled symmetrically via S = max(|r_max|, |r_min|) / 127 and clipped into [-127, 127] for maximum zero-point efficiency.

3

Multi-Object Composite Boards: Synthetic and real-world tarpaulins, street signs, and handwriting combined to train dense multi-label bounding boxes in one pass.

Nano ½ Backbone Stage Configuration 640 × 640 input
STAGE / MODULE REPEATS CHANNELS RESOLUTION
0. Conv 1 32 320 × 320
1. Conv 1 64 160 × 160
2. C3k2 2 128 160 × 160
4. C3k2 2 256 80 × 80
6. C3k2 2 256 40 × 40
8. C3k2 2 512 20 × 20
9. SPPF 1 512 20 × 20
10. C2PSA 2 512 20 × 20
Cultural & Morphological Nuance

The Subtle Scale of "Kudlit" Diacritics

Class-wise analysis in Table IV reveals the primary source of classification degradation: minute diacritic marks altering phonemes between identical base glyphs.

ᜄ
ga (Base)

Base glyph. Precision: 93.5% | Recall: 75.0%

ᜄ᜔
g (Virama Cross)

Cancels vowel. Sub-pixel crosses cause recall drop to 50%.

ᜄᜒ
ge / gi (Top Kudlit)

Top dot alters vowel from /a/ to /i/ or /e/. mAP50-95: 0.532.

Ready to test real-time detection?

Open your camera or upload sample handwriting images. The client-side quantized model performs full inference right inside your browser without backend GPUs.

Start Interactive Demo

Citation (BibTeX) 

@article{nogra2026baybayin,
  title={Real-Time Baybayin Script Detection on Edge Devices Using INT8 Quantized YOLO26},
  author={Nogra, James Arnold and Sambayan, Rachelle},
  journal={University of the Philippines Diliman},
  year={2026},
  publisher={Kaggle Notebooks / arXiv}
}