Basics Machine Vision

Getting Started with Hyperspectral Machine Vision Systems

By Mike Stebbins, COO - January 10, 2024

Hyperspectral machine vision is useful in scenarios that call for automatic inspection, analysis, and action on the factory floor, such as automated sorting. Automated sorters typically use hyperspectral imaging to determine one of two types of information:

Classification: Category prediction (e.g., acceptable vs. not acceptable, product vs. foreign material, specific type of plastic, bruised vs. unbruised fruit).
Regression: Quantity prediction (e.g., ripeness, moisture content, percentage of a material in a mixed group, deviation from a nominal value such as product color).

Hyperspectral machine vision systems collect spectral data from the objects under inspection and then compute the classification or regression on a per-pixel or per-object basis. Depending on the application, the results might show information such as defects in the product, mixture ratios, or debris interspersed among good product.

The results obtained can be translated to meet the input requirements of the factory system, from triggering a subsequent process to alerting an operator to an issue. For example, the system can provide specific coordinates for a picking robot or air jet sorter that will separate the classified products.

Moreover, Resonon’s machine vision software is easy to use. See the video above for a short demonstration. The software’s algorithms learn to detect acceptable and out-of-spec products using training data. The training data are efficiently assembled, and multiple algorithms can be tested for optimal performance. By applying different models, you can determine a solution that meets your application-specific requirements.

How Shape, Texture, Color, and Orientation Impact Hyperspectral Machine Vision System Outputs

Color difference is another way of saying spectral change. Spectral differences, even those too small to be detected by the human eye, are what hyperspectral imaging systems use to differentiate between objects. Object shape statistics (e.g. area, perimeter, orientation) can also be calculated with Resonon’s RVS software, but these statistics do not provide the primary information used by the hyperspectral machine vision system to make predictions.

Product spectral variation and texture generally create challenges for color-sensitive machine vision systems. A hyperspectral machine vision system can overcome this challenge by ensuring that the training data for the classification model include samples encompassing all the spectral variation and texture that the system will encounter in production. In some cases, product texture can even be an asset that improves classification accuracy in a hyperspectral system.