X-ray filter toolkit

Sharpening a blurred foot X-ray with three classic filters, and comparing them with numbers as well as by eye.

P-082026

Project: Sharpening a blurred X-ray, and measuring which method works
Context: Self-directed, built from MSc coursework

48.8 dBbest PSNR, unsharp masking

0.997best SSIM

3filters compared

9unit tests

The question

Which sharpening method best restores a blurred X-ray, and when does sharpening start to create misleading artefacts?

What I did

Modelled blur as convolution with a Gaussian kernel, then sharpened with Laplacian filtering, unsharp masking and high-frequency emphasis.
Ran a controlled test: blurred a clean image by a known amount, sharpened it, and scored the result with PSNR and SSIM inside the anatomy only.
Scored the real blurred scan, which has no clean reference, on edge strength, background noise and clipped pixels.
Packaged the work as a reusable Python module with a scoring script, a figure script and unit tests.

What I found

Unsharp masking gave the best restoration in every controlled test: at mild blur it lifted PSNR from 40.4 dB to 48.8 dB. High-frequency emphasis came second, and the Laplacian mostly added fine-scale noise. Pushing any method too hard creates halos and clipping, which matters because an artefact on a medical image can be read as a real feature.

Figures

The scan as received and after each of the three sharpening methods, with a zoomed view of the mid-foot. — The blurred foot X-ray and the three sharpening methods, with a zoom on the mid-foot.

Chart of controlled test scores for the three sharpening methods. — Controlled test scores for each method.