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
1EduScale/
2├── README.md
3├── models/
4│ ├── eduscale_span_x2.tflite
5│ └── eduscale_span_x3.tflite
6├── benchmarks/
7│ ├── benchmark_x2.csv
8│ ├── benchmark_x3.csv
9│ └── benchmark_summary.csv
10└── docs/
11 ├── MODEL_CARD.md
12 └── 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.
1import tensorflow as tf
2
3interpreter = tf.lite.Interpreter(model_path="models/eduscale_span_x2.tflite")
4interpreter.allocate_tensors()
5
6input_details = interpreter.get_input_details()
7output_details = interpreter.get_output_details()
8
9print(input_details)
10print(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.
