Natural Selection AI AI overview

Artificial Intelligence “NaturalSelection.AI” is essential in automated systems requiring detection of any kind of objects on huge image data. It can be applied to different kinds of industrial products: Textile, Nonvowens, Paper, Glass and many more. The Artificial Intelligence does not produce a computationally expensive Neural Network. It produces an image processing solution, which provides a computationally cheap solution with high accuracy in detecting and segmenting objects for any kind of images.

Minutes

In the cloud

Several GB/s on simple consumer hardware

High segmentation quality

The solution is executed locally – it runs on the hardware of your choice. From small to huge.

Architecture

User Interface

We provide an additional drawing tool for intuitive use within a web browser. No need to install a User Interface Software.

Applications

Line or Matrix Cameras provide material images, the Artificial Intelligence analyzes characteristics of these images and provides a solution for defect and irregularities detection.

Let’s say, our customer wants our Artificial Intelligence to produce a solution for detecting lines close to the car:

1) Load images using web browser.

We want to detect lines close to the car with our Artificial Intelligence.

2) Draw on the images from within the browser:
– in Red the lines you want to detect.
– in Blue what could optionally be detected.
No object should be detected elsewhere.

3) Upload images to cloud and start training AI in the cloud. The training is short.

Artificial Intelligence running in the cloud produces a computationally cheap solution for object detection.

After training, AI produces a solution: objects drawn in Green.

The difficulty of the task lies in the material itself:

– The background contains a lot of noise

– The background displays a wide range of colors and structures

Furthermore, the speed of the material is typically very high, so the solution must be cheap computationally. Following is a small sample of original pictures (top), defects drawn in red (middle) and the computed defects drawn in green (bottom).

The previous examples require the user to mark known types of defects. When production starts, however, this information often does not exist yet.
In this case, unsupervised training is the solution. Only a few images of the undamaged reference material are needed, for example:

NaturalSelection.AI’s artificial intelligence learns during training that this is good material, and that deviations should be considered defects. When real defects occur, they will be detected and correctly segmented, getting production up and running faster.

A few examples with original defects and the resulting segmentation:

Automatic detection and recognition works even for small defects in low-contrast scenarios:

The example images are used with permission of RAM GmbH (ramgmbh.com)