EduScale AI

EduScale AI is a lightweight TensorFlow Lite model release for enhancing educational images such as lecture slides, screenshots, and study material before OCR or visual inspection.

This repository contains two SPAN-based super-resolution models:

Model	Scale	File
EduScale SPAN x2	2x	models/eduscale_span_x2.tflite
EduScale SPAN x3	3x	models/eduscale_span_x3.tflite

Intended Use

• Upscaling low-resolution educational slide images.
• Improving readability before OCR workflows.
• Mobile or edge deployment with TensorFlow Lite.
• Research and portfolio demonstration for education-focused image restoration.

These models are not a replacement for human verification of extracted text. OCR output should still be reviewed when correctness matters.

Data Sources

The education slide data used around this project comes from a mix of:

• Public PPT material gathered from Kaggle: Dataset for ppt by Manisha717.
• Additional educational slide/image samples created by the project author.

The Kaggle dataset page lists its license as unknown, so this repository does not redistribute the original PPT files or source images. Only model files and benchmark CSV results are included here.

Repository Structure

EduScale/
├── README.md
├── models/
│   ├── eduscale_span_x2.tflite
│   └── eduscale_span_x3.tflite
├── benchmarks/
│   ├── benchmark_x2.csv
│   ├── benchmark_x3.csv
│   └── benchmark_summary.csv
└── docs/
    ├── MODEL_CARD.md
    └── DATASET_CARD.md

Benchmark Summary

Benchmarks were computed from the detailed per-image CSV files in benchmarks/.

Model	Scale	PSNR	SSIM	OCR Confidence	CER	Runtime	Device
eduscale_span_x2	2x	29.46	0.9786	91.23	0.0736	573 ms	Realme Note 50
eduscale_span_x3	3x	26.60	0.9642	89.17	0.1980	288 ms	Realme Note 50

Metrics:

• psnr: peak signal-to-noise ratio, higher is better.
• ssim: structural similarity, higher is better.
• ocr_confidence: OCR confidence after super-resolution, higher is better.
• cer: character error rate after super-resolution, lower is better.
• runtime_ms: average runtime in milliseconds on the benchmark device.

Benchmark Files

File	Description
benchmarks/benchmark_x2.csv	Detailed benchmark results for the 2x EduScale model
benchmarks/benchmark_x3.csv	Detailed benchmark results for the 3x EduScale model
benchmarks/benchmark_summary.csv	Short comparison summary of x2 and x3 model performance

Usage

Load the models with a TensorFlow Lite runtime in your target environment. The exact preprocessing and tensor shape should match the model input signature reported by your TFLite interpreter.

import tensorflow as tf

interpreter = tf.lite.Interpreter(model_path="models/eduscale_span_x2.tflite")
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

print(input_details)
print(output_details)

Documentation

• docs/MODEL_CARD.md: model details, intended use, metrics, and limitations.
• docs/DATASET_CARD.md: benchmark dataset notes and CSV schema.

License

This repository is released under the Apache 2.0 license.