Image Background Removal Service (U-Net on AWS)

The problem

E-commerce apparel images need a clean, consistent background, and cutting garments out by hand does not scale past a handful of products. The client wanted a service that takes an apparel photo and returns the same image with the background removed, good enough to use on product pages without manual retouching.

What I built

• A U-Net segmentation model, an encoder-decoder neural network, that separates the garment from everything behind it at pixel level.
• A Python inference service around the model that accepts an image URL, runs the segmentation and returns the background-removed result.
• Deployment of the service on an AWS EC2 instance, exposed over HTTP so it could be called from a browser or from other systems.

How it was delivered

Shipping the model as a hosted endpoint rather than a notebook was the point of the project. On EC2 the client could point the service at any product image URL and see the output immediately, which made it easy to validate quality on real catalogue photos before integrating it anywhere. The original demo instance has since been retired.

Outcome

The U-Net approach produced very good results on apparel photography, and the deployed service on EC2 gave the client a working, demoable background-removal tool rather than a research prototype.