LiıPy Dataset Collection
Crowdsourcing odia handwriting samples through LiPyD.
Model Training
EfficientNetB0 trained on 64×64 images from the Hugging Face dataset.
Downloaded dataset from Hugging Face and filtered valid classes.
Cleaned dataset uploaded to Google Drive.
Drive dataset imported on Colab kernel.
Trained EfficientNetB0 and exported model weights.
System Deployment
Automated hosting architecture built with Vercel and Azure, synced in real-time via GitHub commits.
Inference Pipeline
How handwritten strokes translate to digitized character classifications.
User draws a character on the canvas board or uploads an image.
Grayscales, normalizes pixel intensities, and resizes to 64x64 matrix.
Feature maps computed through Conv2D blocks to output class scores.
FastAPI backend returns prediction results and confidence values.
UI maps label keys to Odia glyphs and displays predictions to user.
Technical Challenges
Key script obstacles encountered when training neural models on handwritten Odia data.
Different writers produce characters with varying stroke sizes, shapes, and speeds.
Handwritten Odia research lacks large, structured, and standardized repositories.
Visual overlaps in character curves make accurate classification highly difficult.
System Metrics
By the numbers: character class distributions and dataset properties.
Includes Odia independent vowels and consonant characters.
Crowdsourced samples collected from active contributors via LiPyD canvas.
EfficientNetB0 trained on augmented handwritten samples.
Operational web interface connected to cloud REST backend endpoints.
Technology Stack
Core frameworks and software architecture powering the platform.
Next.js frontend deployed on Vercel with canvas input and result dashboard.
FastAPI inference service deployed on Azure with Hugging Face model loading.
EfficientNetB0 model loaded from Hugging Face Hub via TensorFlow/Keras.
Frontend on Vercel, backend on Azure, datasets on Hugging Face, data on Supabase.
Roadmap
Future milestones and feature roadmap for the LiPy ecosystem.
Adding complex compound character classes to cover more ligature curves.
Shrinking model weights and quantizing parameters for low-end mobile edges.
Building line segmentation modules to scan whole paragraphs instead of characters.
Exposing REST endpoints to academic researchers for script recognition.