DEEP LEARNING IN INDUSTRIAL INSPECTION
Can DEEP LEARNING be used as a universal tool for machine vision, just like a Swiss Army Knife? We at AIT work at the forefront of machine learning and we know when deep learning is the appropriate tool for solving your problem and when it’s not.
ADVANTAGES OF DEEP LEARNING
Deep learning can minimize that costly and labor-intensive process and at the same time increase speed and accuracy of quality inspection.
Deep Learning methods like CNNs (Convolutional Neural Networks) are methods that imitate the principles of the human brain to solve problems. CNNs are trained on a specific set of data. By analyzing a vast database, the software can spot and recognize patterns. Thus, in contrast to classical methods in which models are designed and cast into formulas, even blurred problems can be solved.
DEEP LEARNING CAN BE USED FOR
With the help of deep neural networks, one is able to solve various tasks in the field of industrial inspection, e.g.
- defect classification,
- error and anomaly detection, and
- substitution of complex imaging algorithms for better quality and runtime.
INDUSTRIAL CHALLENGES FOR DEEP LEARNING
Training still requires a huge amount of annotated data. In industrial production processes, there is a lot of data about what a flawless product should look like, but there are not enough images of defects, which are as important for a successful training of a neural network as good samples.
With the experience from our research conducted in special development environments we can:
- Keep deep learning solutions as simple as possible.
- Ensure deployability and high processing speed.
- Minimize the effort for deployment on different platforms.
- Investigate available deep learning deployment Hardware technologies.
- Use our own deep learning solutions to deploy them on standard hardware for deployment verification.
DEEP LEARNING CAN DO A LOT FOR YOU...
Deep learning methods are powerful and appealing. By using deep learning, it is possible to solve problems and tasks where traditional image processing algorithms fail.
- Substituion of Complex Algorithms
- One Class Learning / Anomaly Detection
- Data Augmentation
... AND WE KNOW HOW TO USE IT.