An overview of how we use Machine Learning to edit & create new images at FlixStock
About 50% of any eCommerce website is full of images! And it is this 50% that is the most crucial factor in influencing the decision of most buyers.
Yet, 99% of images are created and edited by skilled ‘Humans’ — a repetitive & mundane job. At FlixStock, we work on replacing this repetitive jobs with advanced software solutions.
Creating & editing images using Artificial Intelligence helps in:
- bringing down the error rate drastically,
- increase the speed of image editing & most importantly,
- let the skilled ‘Humans’ use their brilliant mind in marketing & creative images!
Machine Learning & Image Editing
So how can Machine Learning help? At FlixStock, we create model images without using an actual model! All we need is the picture of a garment to do this.
How do we do it?… I have visualized the process we follow, below:
This is a constantly evolving process with at least a dozen Machine Learning & Deep Learning programs running simultaneously and/or sequentially. We have a couple of Research Scientists working round the clock to make this process more efficient than ever.
This article is an initiative to share our approach, experiences & learning with the growing AI community.
Deep Learning & Image Editing
While we have seen the successful implementation of Deep Learning (DL) in data analytics to predict your favourite show or identify objects in an image, DL is still not widely used in the process of ‘creating’ images. Image Processing using Machine Learning is another component.
It is interesting how this ‘more-than-ever’ popular form of content (read images) — which occupies half of the world’s internet traffic, is not a priority in our field. Though there have been many fun projects & some very serious ones, we are yet to come across a large scale implementation of Deep Learning in Image Editing.
Time for a fun break: I came across this interesting project where DL was used in Image Creation. All you had to do is draw a cat & the machine will make a ‘cat out of it’! A gif along with the link to more results, below:
I admit that most of the cat images generated by #Pix2Pix experiment (http://bit.ly/2OQS6qW) are far from perfect, but most do definitely resemble a cat! It reaffirms that images can indeed be generated with AI.
Semantic Segmentation & the art of clipping!
For extraction of an identified object (for e.g.: a human model from studio background), one of our research scientist (Beeren Sahu) chose Semantic Segmentation after many trials. For the AI enthusiast in you, you can read the details of his experiments with Semantic Segmentation here: click here
With a database of few hundred thousand images, we were able to generate a near perfect image of the extracted object. Like in the case of this batch of t-shirts we processed through our solution.
However, there were two main challenges:
- For faster processing, we were forced to use small (size) images.
- Transfer learning would have been faster, however, the quality of output images required us to start from scratch.
Resulting images, for a trained eye, had issues at pixel level along the edges. Naturally, Semantic Segmentation (or Deep Learning) has its limitations. But, with every new dataset, we are able to narrow down on this issue.
For practical purposes, we integrated Image Processing (ML) along with Semantic Segmentation to make the images better & usable.
Not the desired solution as it increases the time required & is a huge hogger of processing power. However, its a quick fix till our algorithms are better trained (in near future).
Now that we have shared our recipe for extracting the images through Semantic Segmentation, our research team is tasked with identifying the correct model pose/stance & accessories for a garment— a task which sounds very simple, but essentially herculean. We will be sharing our experiences in tackling the same in a subsequent article, shortly.
Meanwhile, we could use your encouragement in the form of ‘like’ & ‘share’ on this article.