A smart solution that would automatically detect and remove the background from pictures of people and objects.
US | Tech Industry
An American company from the tech industry.
A global brand with headquarters in Australia, established almost 20 years ago. Their offices on few continents hire more than 500 people and provide services to a couple of millions of clients every year.
This company wanted to create a smart solution that would automatically detect and remove the background from pictures of people and objects. There are many obstacles that we could stumble upon while building this type of product. For example, we needed to program the machine with attention to the detail – it is hard to remove background from pictures of people with messy hair or small clothing details on the borders.
We’ve picked Google Cloud for training the machine learning model and conducting experiments. Using a GPU-based infrastructure for training the machine learning model and conducting experiments was a must in this case. We used Google Cloud servers because of the platform’s scaling ability that was sufficient for the amount of data that needed to be processed. The outcome was a machine learning network model dedicated to clearing backgrounds from images of people or objects.
It was a perfect opportunity to use machine learning while developing this product. It is worth mentioning that every machine learning project is a collaboration between programming and computer science. We need to work with each other to achieve results that will meet the expectations of the client.
Experts we were cooperating with provided us with the architecture proposal based on the current state of the art conducted from scientific data.
At first, we’ve determined an initial architecture for a neural network. Then we would move to the next phases of the project. We’ve set a direction for learning experiments to be performed – with defining the criteria of precision measuring and determining what kind of data will be needed to teach the machine to remove backgrounds from pictures.
In a machine learning project like this, you always start with small cases of learning examples and then you progress with something more complicated. For instance, to teach the machine about contrasts, we can use a step-by-step method and provide the computer with pictures of one person on a simple background with light skin and dark hair. That won’t be as challenging as group pictures on complicated nature-based backgrounds like rocks or trees. We can move to the next phase of recognition when the first one is completed and works as it is supposed to.