Computer vision in the industrial sector. How to improve automated quality control?

Computer vision has already paved its way to the mainstream and entrenched its position as one of the most universal and “industry-agnostic” technologies in the AI field. Thus, using it to streamline quality control processes is a no-brainer.

Long story short, it is about drone-based quality control. A global company – whose name we cannot disclose for now – manufacturing and selling components for all vehicles teamed up with us to develop an innovative quality control system based on drone-based computer vision, precisely, a system based on drones equipped with RGB and IR cameras.

The company wanted to control the quality of the idlers transported on the conveyors. As the conveyors were very long and localized in hard-to-explore terrain, manual quality controlling these objects was challenging or even – at some places – impossible, no mention that its accuracy was very disputable.

Replacing it with automated solutions was a natural direction. The idea was straightforward: drones flying over the conveyor belts were supposed to detect the idlers that needed to be replaced by analyzing their temperature. The prototype was already done and even worked. The problem was it worked poorly – it was slow and sometimes sloppy. The client was aware of these issues and had an idea of how to improve them.

The client was deeply aware that it is data that determines the quality of the Machine Learning model and – which is by no means a typical case – provided a considerable amount of well-structured data to “feed” the model we were about to deliver. – said Michał Myller, Data Scientist at Addepto.

Addepto’s main task was to migrate the system to technologies that would improve the system’s image detection speed while increasing the process’s accuracy. As the prototype already existed, the first month of work the Addepto team spent on due diligence to find out what elements required taking action and what worked just fine. After this time, the team delivered a report with suggested improvements, which the client accepted. Then, the time for the implementation started.

The key is time and accuracy in processing data

As the computer vision area is well-explored in AI fields, the Addepto team had no intention of reinventing the wheel. Instead, it reached out for state-of-the-art solutions to deliver a smoothly working platform with a user-friendly UI.

The delivered solution was based on the YOLO framework for object detection, deploying MLflow in AWS, using SageMaker Pipelines for automation and model versioning, and CodeCommit for Docker builds. These technologies gave the developers certainty that the app performance would increase.

Computer vision state-of-art technologies

The client had a vast amount of human-annotated data ready to use in the model, but – to make the most out of them – the Addepto team checked and corrected its structure and then divided it into two sets: training and testing to establish a clear benchmark to compare future solutions.

Then, there was speed improvement left. The client’s solution needed around 20 seconds per image to process one image running on the CPU. Thanks to the technologies Addepto used – lightning-fast YOLO, GPU inference, and Sagemaker’s inference frameworks YOLO – the app’s overall performance significantly increased.

Thanks to implemented improvements, the time needed to process the data went from 30 minutes to ~8 minutes per ~200 high-resolution images. Also, by creating the testing and training data sets, Addepto could check its system’s accuracy – all it needed to do was compare the results. Addepto delivered the system that fit the Client’s requirements – fast and accurately.