In 2014 one of professional photo labs contacted us to help them develop an image processing system.
The system was intended to automatically remove a solid color background from photos and correct the problematic issues such as improper exposure, shadows in background or improper lightning setup.
A large digital photo services company.Industry
Professional photography studios
and color–processing labs.
The previous system in place removed the background from the photos, but could not handle a number of effects pertinent to digital photography, such as chromatic aberration, blooming, micro-motion blurs, etc., and also effects due to bad lens quality. As a result, the system produced images with lots of false transparent areas. The company had to extend its staff to manually correct processing mistakes using tools like Adobe Photoshop.
From the business standpoint, this led to a higher cost per image, longer learning curve for the new staff, and made it difficult to scale up and down, taking into account the seasonal nature of our customer’s business.
Our team found and reviewed multiple computer vision algorithms. We chose the ones that worked best, and built onto them to create an elaborate system, capable of «recognizing» what’s on the images, and what’s more — learning from them.
The system is using a «deep learning» approach, which is based on training multi-layered artificial neural networks. The system is «trained» statistically, by providing it with a great deal of real-life examples for analysis.
Since the recognition mechanism still is not expected to be 100% perfect, the system encloses tools for fast corrections of remaining processing defects. All such corrections are used as feedback for further training of the system.
In its current state, the neural network behind the updated customer system contains about 15 million parameters. Its training from scratch would take a week, and require thousands of examples of effects it should learn to recognize.
Deep learning has helped us to conquer the undesired effects on photos, which usually come in twos or threes on the same image. By feeding the system statistics on such cases, we trained the algorithm to recognize such effects and account for them during removal of a background on photos.
In this project, deep learning has helped to reduce the amount of manual interference required for the automatically processed images. Up to 85% of resulting photos do not need further correction.
The developed system provides a number of unique capabilities, such as:
Replacing green and blue background of various quality and texture with a transparent one.
Differentiating the background from objects of the same color, e.g. a green person’s shirt on a green background.
Removing the undesired effects common to digital photography, such as micro-motion blurs, oversharpened edges, chromatic aberrations, blooming etc.
Easy-to-use retouch system that allows to fix possible issues of the automatic processing quickly, thus maintaining high order turnaround time.