Deep Learning Vision Systems

Examples of applications with the use of AI

The surface inspection of a curved housing is one of the most problematic tasks in conventional machine vision. When used with AI, a wide variety of NOK examples are labelled, taught and reliably found in the evaluation.

Evaluation result of the determined housing defects. 

Character recognition, even of embossed characters cropped in the image – 99% recognition.

Type recognition based on a taught-in feature and completeness check of sealing rings – 99% recognition.

AI technology for evaluating image data uses neural networks to achieve a learned knowledge that allows it to distinguish between anomalies, shapes and characters while tolerating natural deviations. Thus, artificial intelligence combines the superior flexibility of humans with the performance of a machine system.

How is an AI machine vision system created?

To prepare a Deep Learning system for use to reliably evaluate images, it initially needs a sufficiently large and easily identifiable training set of images with which the artificial intelligence can acquire its knowledge.

To learn the features (classification) that are important for the inspection task, so-called labels (rectangular bounding boxes) are needed, which are generated either automatically or manually in the image. Labelling tools help the user to create the classes and position the labels.

The current state of hardware allows Deep Learning algorithms to handle a mass of data from high-resolution images to easily handle standard image processing applications such as classification, object detection and segmentation. For all areas, there are object recognition pre-trained networks that help to shorten the training time of a well-functioning model. A trained model then corresponds to the learned knowledge that can be drawn on during later evaluation by the AI.

How many images does a training set need to contain?

This depends on many factors, such as the complexity of the inspection task, the amount of possible deviations and the size of the image data. For a simple inspection task, it may be sufficient to train with 25-50 images. However, as the complexity of the task increases, this may well extend to 500-1000 images and more. A variance of the images can also be created simply by artificially manipulating the image data, e.g. by twisting, distorting or changing the brightness.