This is some text inside of a div block.
 min read

Fully Automated Image Cleanup

DRiV (former Federal-Mogul), a publicly traded car parts manufacturer, produces more than 20 new car parts every day. These car parts are photographed from various angles, touched up in photoshop and put in a catalogue. This process is repetitive and time consuming.

Goal

Automatically process the images and remove the background of the photographs.

Impact

The AI product got deployed in production and hit the following metrics:

prophecy labs infographic

Cleaning pixel accuracy: 99%

prophecy labs infographic

Man hours saved per year: 11,000

prophecy labs infographic

Time to clean an image: 0,7s

Problems

  1. Binary semantic image segmentation on high resolution (6000 x 4000) images
  2. Image cleaning quality assessment.

Solution

CLI that cleans a directory of images, in batches on the GPU, while staying true to the folder structure. A machine learning model isused to make the background of the images transparent, holes in the images are filled and car part edges are smoothed. After, another machine learning model assess the cleaning quality and images that are not sufficiently clean are send for a manuel touch-up.

Technical Highlights

Researched and developed a custom deep learning architecture to perform semantic image segmentation based on UNET and PSPNet.

Developed a novel approach to deal with jagged edges inherent to background removal. 

Gathered user feedback regarding the cleaning quality through a Tinder-like interface, to develop a custom deep learning model that could predict how well an image was cleaned.