Our client is an agency specialized in high-volume image processing. They already had theirown image processing algorithms for extracting images from backgrounds but it didn’t work for all types of backdrops they wanted to operate with.That is why they came to us with ...
Our client is an agency specialized in high-volume image processing. They already had theirown image processing algorithms for extracting images from backgrounds but it didn’t work for all types of backdrops they wanted to operate with.That is why they came to us with a clear task, to develop a more suitable solution.
The main application area of the project is photo processing. One of the most common andsought-after
tasks here is to replace the original backdrop of an image due to the need eitherto fix some issues with a background or to make the image more impressive and unique ingeneral. It is also one of the most time-consuming and tedious tasks in the area ofphotographic processing.
Doing this process manually will take an average specialist significantly more time than doing it automatically with an algorithm.
There’s plenty of solutions designed to help specialists with that task on the market built with Computer Vision mathematical algorithms and Machine Learning approach. However, the existing solutions proved to be either too slow for the real production needs or not precise enough.
Challenges
Within this project, we've tested several state-of-the-art neural network architectures in application to the problem of accurate segmentation and computing the transparency levels for border pixels. This work is intended for processing photos with people captured in full length.
Solutions
We've overcome several challenges connected with optimaldata preprocessing necessary for training the network andprocessing the high-resolution images keeping their originalsize by making and aggregating predictions for overlappingpatches.
Results
The project comes up with the solution which helps photomontage specialists to handle high-volumetasks with the required speed and precision where 4000 x 3000-pixel images can be processed inroughly 5 seconds. The resulting solution operates with the following types of backdrops:
- natural,
- painted,
- lastolite.
Business value
A twofold reduction in time spent by designers on photo processing;
The payback time for the project was 2 months.
