Bringing machine learning to steel surface quality inspection

Can a software actively learn to detect and classify defects?

The Tata Steel challenge

Tata Steel came to Scyfer with the challenge of improving their surface inspection process. This is a crucial step in their quality process – even a small oversight in detecting a defect can easily break machinery in operation. The risk involved is huge.

Their system had to classify millions of images a day but fell short when it came to detecting visually complicated defects. As a result, combining human expertise with software capacity was essential to their process. However, when you deal with an astounding amount of data on a daily basis, the support offered by the system needed to be better.

How did Scyfer help?

Scyfer sought out to properly understand how the expert’s eye works in detecting the visually complex defects and how they deal with rare or less known defects. Next up, we built a custom-made machine learning platform that uses computer vision in order to detect the defects.


steel plates surface inspection


The images are produced every day, so that helps in building, training and testing the model. The software becomes better and better at its work the more examples we feed into it. We call this “active learning”.

Real business impact

The solution Scyfer built for Tata Steel impacts the company’s bottom line through improving some key processes for the company. Thus, the benefits of the Scyfer’s platform are:

  • Far less dataset management
  • Significantly higher performance of the defect classification system
  • Successfully capturing the knowledge and expertise of people using it
  • The software improves over time the more data it receives.

As the software learns to better recognise the steel surface issues we aim to achieve a 100% perfect surface inspection for fully automatic coil release.


Steel plates TATA

More on active learning

Active learning can have a positive impact when:

  • You have a lot of data – either a dataset or a continuous flow of data coming in.
  • This data is checked/assessed by humans
  • The outcome of the assessment is critical to operation
  • The amount of data to be assessed is too big to handle, or you can only assess a small subset but you want to scale up.

It works with algorithms that detects objects and classifies them. This works best as a way of automating a visual process (traffic detection, image tagging, cancer detection etc). The benefits of an active learning platform include:

  • Pooling expert knowledge of a lot of people at once
  • Easily integrates into an organization’s workflow
  • Delivers state of the art image recognition results
  • Requires zero hardware changes
  • The solution can be securely accessed by connecting to the Scyfer cloud.