Regression modeling is used to estimate a numerical value instead of making a decision. It is often adopted in quality estimation.
perClass Mira implements object-centered regression modelling. Individual objects can be annotated with numerical meta-data. A regression model is build that can be applied to a new object.
This enables practical applications such as:
- Estimate brix (sugar) content in a tomato
- Estimate a dry-matter content of a plant leaf
- Estimate moisture content in biscuits
Regression example: Powder mixture
In this example, we will use a mixture of sodium carbonate and flour. A set of vials with different mixing proportions is scanned using Specim FX17 camera in the range between 900 and 1700nm.
In order to estimate mixing proportion we:
- Build a classification model that identifies area of interest for regression
- Define the class of interest as foreground so that objects can be segmented out
- Annotate number of objects with known ground-truth values for regression
- Build a regression model
- Flag some of the images for testing only to judge generalization performance
- Improve the model
- Apply the solution to a new hyperspectral scan (objects get identified and for each a value is estimated)