Automatic, algorithm-based image recognition

Purpose and goal

The objective of our project was to make it easier and more scalable to sell large volumes of images both for a stock image agency and for photographers, through creating image and object recognition software that drastically increases the efficiency of image publishing. With our project we have accomplished an important step towards this goal, which is to automate the review of whether or not a model contract is required to sell an image.

Results and expected effects

Mostphotos is on track to become the first company to use image recognition technology for image review of model contract requirements. This is an important milestone towards our goal to simplify selling images for both hobby and professional photographers. It implies significant cost savings as costs for manual review decreases and also much greater possibilities to accept larger volumes of images to our image archive. We also do not exclude the possibility to sub-license our technology.

Approach and implementation

We have used the Convolutional Neural Networks (CNN) model for classification, localization and detection of objects on images. Furthermore, we´ve used deep learning to optimize the system´s ability to localize objects and recognize object boundaries. CNN´s have proven to be superior, especially when it comes to object classification, where CNN´s has outperformed even the human eye. Additionally, we have used a few hundred thousand images from Mostphotos´ archive to train and test the system on.