Image Retrieval by Deep Learning

Using Convolutional Neural Networks to find similar images among millions of photos

The marketing companies, sending people on the field, require as a confirmation a picture taken in or in front of the premise. Therefore arises the need to check if a picture from the same place, taken at the same time, or a transformation of it (crop, resize, blur, brighten), has already been sent or not.

Because of this we created a product which uses machine learning – CNN (Convolutional Neural Networks) to create appropriate image representations. With an adequate training and choice of function, our solution creates a metric which evaluates the proximity of the two pictures based on their portrayal. On the other hand this same metric gives you the option to index all of the pictures speeding up the searching dramatically. For the sake of productivity the representations are being converted to ones with minimum size without losing almost any accuracy. The end result is the ability to find similar pictures among a million other photos with more than 90 % accuracy with a minimum amount of false positives for the staggering 0.2 seconds.

The product idea was proposed by a client of ours – the number one Mystery Shopper software producer in the world. The company is already integrating our solution into their system. Finding similar images in a data set can also be applicable in other business areas.