A professional photographer spends hours making subjective qualitative enhancements to each photograph.

Meero assigns machines to learn how to do it in order to offer a fast, global, and scalable solution.

Dedicating artificial intelligence to photography

In order to deliver tens of thousands of photographs per day, developing an automatic image enhancement pipeline has always been a crucial issue for Meero. Based on the latest advances in machine learning research, it has become possible to program touch-ups using cutting-edge techniques.

The most important problem to solve consists of developing an algorithm that would perform the same global and local image transforms as a professional retoucher. Meero's expertise in the field, combined with a unique 25-million photograph database, has built a strong foundation for this machine learning paradigm


Learning to enhance photographs

Meero uses a dual-deep convolutional neural network involving global and local feature extraction on a multi-scale basis, in order to make the algorithm thoroughly learn from the manually enhanced photographs, and to make it robust to changes in lighting (including white balance and color grading), scene semantics, and capture defects (such as noise or motion blur) as well.

Extracting valuable insights from photographs

Beyond the aesthetics of a photograph, Meero’s AI is also able to turn the image’s visual attributes into meaningful and valuable insights. The power of deep learning can once more be put to good use. The assets of semantic annotations also available in Meero’s database, are used besides the photograph enhancements. By coupling pictures with metadata including bounding boxes as well as class indexes, it is then possible to predict the presence of an object in an image, in a fraction of a second.


A cutting-edge algorithm inspired by
how the human brain works

How is such a thing even possible? How can an algorithm perform as well as a professional photographer when executing such a subjective and complex task? The answer resides in the principle of gradient-based learning. By using a deep neural network that mimics the way the brain works, Meero can make use of the errors the network’s neurons output and compute its gradient using convex optimization methods. By then, using the chain-rule to correct this error from the end to the beginning of the network, and repeating the operation millions of times using Meero’s database, the proprietary deep learning algorithm performs as well as a professional photographer when enhancing pictures.

Convolutional Neural Network