Automatic Visual Inspection

Heavy industry

Background

The manufacturing company is producing precise, high-quality electronic components. Due to increasing, production volume defects have to be detected faster and more precisely than ever before. Strict quality requirements had to be met on all of the stages.
The company wanted a video inspection system to monitor the semi-finished products during the production process and recognize the broken ones. This work was needed to be automated because precision-time trade-off in the case of humans was unacceptable.

Project Goals

  • Design infrastructure for gathering and processing the data online and on spot.
  • Develop and deploy a real-time model for visual inspection.
  • Integrate the system with an existing production line and determine new business processes.

Challenges

  • The sizes of the defects were varied.
  • Due to the size of elements, high quality and detailed images are a must.
  • Response time has to be at least as fast as the throughput of the production.

Our work

The audit of the production line with respect to possible infrastructure installation was held. A few possible cameras setups were tested to make sure the quality of data will be good enough. We designed a data pipeline to acquire the video signal from the cameras, preprocess it and provide it to the model.

The main idea was to extract single frames from the videos and then use the convolutional neural network to segment the objects of interest from the images. Then the second deep learning model was trained to detect and mark any defects on products’ surfaces. If the defective element was found the information was logged in the database and further steps were held to take an element out of the production line.

The outcome

Afterward, when results were acceptable for the company the whole infrastructure was designed and installed in the production line and necessary integrations were done.

Benefits

The company was able to reduce the number of complaints.

  • The defects recognized on the early stage of the production process significantly reduce the number of malfunctioning end products – savings on wasted non-defective elements.
  • Production becomes scalable thanks to fast response time.
  • Responsibility was taken of the employees.
Technology used
  • Python
  • TensorFlow
  • OpenCV