Architecture
10,634 params · DW-Sep Conv · GAP
🧠MyanNet — Layer-by-Layer
Block 1Conv2D(64,3×3) + BN + ReLU + MaxPool
Block 2DWConv(3×3) + BN + PW(1×1,64) + BN + MaxPool
HeadGAP → Dropout(0.194) → Dense(64) → Softmax(10)
Total trainable params10,634
DW MACs reduction~8–9× vs standard 3×3 Conv
GAP head vs FlattenO(C·N) vs O(H·W·C·N)
Results
BHDD Test Set · 27,561 samples
⚡ Edge Deployment
TFLite model size24.18 KB
Mean latency0.263 ms/image
P50 latency0.262 ms/image
P95 latency0.299 ms/image
Benchmark runs1,000
Throughput>3,800 img/sec
ARM Cortex-M SRAMFits without ext. flash
📊 Model Comparison
Baseline CNN34,826 params · 99.58%
GAP-BN CNN21,418 params · 99.51%
MyanNet (ours)10,634 · 99.49%
CNN (myMNIST)34,826 · 99.70%
PETNN (GELU)N/A · 99.66%
Improved CNN [BHDD]~431K · 99.83%
Param reduction69.5% vs Baseline
🔁 5-Fold Cross-Validation
Mean ± Std
99.46% ± 0.06%
📈 Training Dynamics
Train Acc
Val Acc
Loss
Acc surpasses 99% within 5 epochs. EarlyStopping at ~epoch 28. Two LR reduction dips visible on val curve.
🔍 Optuna — 30 Trials (TPE Sampler)
Best Hyperparameters
filters1 / filters264 / 64
dense_units64
dropout0.1937
learning_rate0.003005
🔄 Data Augmentation
Rotation range±15°
Width / Height shift10%
Zoom / Shear range10%
Applied duringTraining only
🎯 Confusion Matrix — BHDD Test Set (27,561 samples · 99.49% accuracy)
Errors concentrate on class 0 (68 / 6,856 misclassified) → class 1 (20 errors, script-intrinsic: closed circle vs open arc) and class 6 (12 errors).
Class 1 confused with class 5 in 17 cases (glyph similarity). Classes 6–9 achieve near-perfect recognition (389–610 test samples each).
Conference Paper
IEEE Format · 2026
MyanNet: A Lightweight Convolutional Neural Network with Depthwise Separable Convolutions
for Burmese Handwritten Digit Recognition
Ah Maung Oo
Dept. of Computer Science & Engineering
Marwadi University, Rajkot, Gujarat, India
ahmaungoo24@gmail.com
Dr. Tarannum Bloch
Dept. of Computer Science & Engineering
Marwadi University, Rajkot, Gujarat, India
tarannumbloch@marwadiuniversity.ac.in
Abstract
Recognition of handwritten digits in Burmese script remains underexplored in the OCR literature, particularly under the constraint of inexpensive hardware.
Existing work on the Burmese Handwritten Digit Dataset (BHDD) optimizes solely for classification accuracy, neglecting model size and computational cost.
We present MyanNet, a compact CNN combining depthwise separable convolutions, global average pooling (GAP), and batch normalization,
designed to approach state-of-the-art accuracy within a parameter budget suitable for mobile and embedded deployment.
MyanNet achieves 99.49% accuracy on the BHDD test set with only 10,634 trainable parameters—
a 69.5% reduction over a standard CNN baseline at a cost of 0.09 percentage points.
A 30-trial Optuna search identifies optimal hyperparameters; 5-fold cross-validation confirms stability (99.46% ± 0.06%).
After TFLite integer quantization, the model occupies 24.18 KB and runs inference in 0.263 ms on CPU.
To our knowledge, this is the first BHDD study to treat edge deployability as a primary design requirement.
Burmese digit recognitionMyanmar OCRLightweight CNN
Depthwise separable convolutionGlobal average pooling
OptunaTFLiteBHDDEdge deployment
🎯 Contributions
① MyanNet: 10,634-param CNN, 99.49% on BHDD
② 30-trial Optuna search (TPE sampler)
③ 5-fold CV: 99.46% ± 0.06%
④ TFLite: 24.18 KB · 0.263 ms/image CPU
⑤ First BHDD edge-deployment study
📉 Efficiency Analysis
vs Baseline CNN (params)−69.5%
vs BHDD best (params)40× fewer
Accuracy cost vs baseline−0.09 pp (~25 errors)
DW conv computation~8–9× reduction
GAP head paramsO(C·N) vs O(HWC·N)
Fits ARM Cortex-M SRAMYes (no ext. flash)
🔬 Training Setup
FrameworkKeras 3 / TF 2.x
Random seed42 (Python/NumPy/TF)
LossCE + 0.1 label smoothing
Batch size64 (MirroredStrategy)
LR scheduleReduceLROnPlateau ÷2
Early stoppingpatience=10, ~epoch 28
Max epochs100