Real-Estate Image QC

Real-estate photography goes through editing at volume — exposure correction, HDR blending, color work — and quality control is usually a person comparing the edited image against the original. That is slow, inconsistent and hard to scale across thousands of images.

The question was whether image-quality signals could flag likely failures automatically: over- or under-exposed edits, color shifts, HDR artefacts, or edits that drifted too far from the reference. Before any model is trained, someone has to prove which signals actually separate good edits from bad ones.

I treated this as an experimentation and QC problem, not a model-training project. The first job was a clean dataset: collecting edited and reference pairs, preprocessing them consistently and organising them so every experiment ran on the same inputs.

From there I worked signal by signal — scene classification to route images to the right checks, image comparison and similarity to measure drift from the reference, and mask-based analysis to focus on the regions that matter, such as windows and skies where HDR problems show up.

The workflows were designed as reproducible notebooks: each experiment loads the same prepared dataset, computes one family of signals and writes results as tables that can be compared across runs.

Outputs were kept visual wherever possible — side-by-side comparisons, difference maps and masked regions — so failure cases could be inspected by eye and used to refine the QC rules.

Preprocessing, masks and image-quality analysis were built with Python and OpenCV; PyTorch handled scene classification and image-similarity features; NumPy and Pandas managed the numeric signals and result tables. Experiments ran in Jupyter locally and on Google Colab when they needed more compute.

Image-quality checks covered color and exposure statistics, similarity between edited and reference images, and HDR/enhancement differences — the signals that most often explain why an edit fails review.

This project had no public web surface, so there was no search work. Its value is in the data and the QC logic it produced.

Benchmarking was part of the work: measuring how each signal performs at separating passing from failing images, and how long each check takes, so the QC logic stays fast enough to run across large batches.

The experiments produced a prepared dataset, a set of reproducible evaluation notebooks and a clear picture of which image-quality signals identify failure cases — the groundwork for AI-assisted QC logic in a real-estate photo pipeline.

It is also where my web and AI work meet: the same discipline of measurable checks that I apply to Core Web Vitals and SEO, applied to images instead of pages.