Lê Chí Đại
← Projects
Image processing · Classical methods · 2025

Image Super-Resolution

A modular image restoration project comparing three families of traditional super-resolution algorithms, without deep learning.

RoleProject leader · Team of 5
Year2025
ToolsNumPy · OpenCV · Gradio

Project visuals

Figures from project reports, pitch decks, and saved experiment outputs. Select an image to open it at full resolution.

Three panels comparing predicted high-resolution output, low-resolution input, and the high-resolution reference of a bird.
Patch-based reconstruction: predicted HR, LR input, and HR reference. Saved output from the team’s experiment notebook (Set5 bird example). Source ↗
Four bird images comparing upscaled low-resolution input, bicubic interpolation, A-plus reconstruction, and high-resolution reference.
A+ experiment: LR upsampled, bicubic, A+, and HR reference, with the notebook’s reported PSNR. This is a team experiment, not the iterative back-projection result. Source ↗

The problem

Upscaling an image does not recover information automatically. The project examines how interpolation, reconstruction, and example-based techniques behave under a shared evaluation setup.

My contribution

Led a team of five across three algorithmic tracks and coordinated integration. Implemented an iterative back-projection engine with a vectorized spatial refinement loop in NumPy.

Results

The iterative back-projection implementation converged within 15 iterations, with a reported +1.85 dB PSNR gain over a bilinear baseline across Set5 and Set14 experiments. The project also includes a Gradio interface for exploring methods and evaluation metrics.

What matters beyond the score

PSNR captures pixel-wise error, not the entire perceptual experience. Edge sharpness, ringing, texture artifacts, and assumptions in the degradation model deserve inspection alongside quantitative metrics.

Next project

ML Benchmark Suite