Dry beans

Sort and classify beans using features extracted from images.

Motivation

The common bean (Phaseolus vulgaris) is a plant widely cultivated for its dry seeds—beans. The species has long been cultivated for food and has developed into many varieties; many common beans, such as kidney beans, pinto beans, navy beans, and wax beans, are all varieties of the common bean. Because there are many varieties within the species, there are many varieties of seed that farmers can choose to plant, and each variety is suited for different growing climates and for different markets.

However, in countries like Turkey, most farmers do not use commercially produced seeds certified to be all of one specific variety. They use seeds they obtain from their own crops or from other farmers, which may contain many different varieties of bean. To sort out specific varieties would require a trained expert to carefully sort the beans—which would be quite difficult when planting a field takes many thousands of seeds!

As a result, there is some demand for systems that can automatically classify seeds by variety, for example by using images of the seeds. If seeds can be accurately sorted by computer, a machine could automatically sort seeds at high speed, allowing farmers to sort their crops, pick the best varieties for their fields and their markets, and sell their seeds to other farmers who want a specific variety.

To collect this data, researchers used a camera to take photographs of 13,611 beans of seven different varieties. (This study was conducted in Turkey, so the varieties are known by their Turkish names: Barbunya, Battal, Bombay, Calı, Dermason, Horoz, Tombul, Selanik and Seker.) The beans were of known variety. The researchers then used image processing code to extract features from the images, including the bean sizes and several measures of shape. The goal was to use these features to classify the variety of the beans; if this is possible, one could build a system using a camera and computer to automatically classify beans by variety.

Data

This dataset contains features from 13,611 beans.

Data preview

dry-beans.csv

Variable descriptions

Variable	Description
Area	Area of the bean in the image, measured as the number of pixels it covers in the image
Perimeter	The circumference of the bean, measured as the length of its border in pixels
MajorAxisLength	The length of the bean’s major axis in pixels, i.e. the longest line that can be drawn from end to end of the bean
MinorAxisLength	The length of the bean’s minor axis in pixels, i.e. the longest line, perpendicular to the major axis, that can be drawn across the bean
AspectRation	The ratio of major axis to minor axis length
Eccentricity	Eccentricity of an ellipse with the same second moments of area as the bean
ConvexArea	Number of pixels in the smallest convex polygon that contains the bean
EquivDiameter	The diameter of the circle with the same area of the bean
Extent	Ratio of the area of the bean to the area of the smallest rectangle that contains it
Solidity	The ratio of the bean’s area to its convex area (i.e. Area/ConvexArea)
roundness	4πA / P², where A is the bean area and P its perimeter
Compactness	Ed/L, where Ed is the EquivDiameter and L is the major axis length
ShapeFactor1	Major axis length divided by area
ShapeFactor2	Minor axis length divided by area
ShapeFactor3	4A/(πL²), where A is area and L is the major axis length
ShapeFactor4	4A/(πLl), where A is area, L is major axis length, and l is minor axis length
Class	Variety of the bean

Questions

1. Using a multiclass classifier—such as a neural network, multiclass support vector machine, or kNN—build a classifier to predict bean variety from the available features. Choose the features you use carefully; because some features are calculated from others, some types of classifier may not need all features. Also choose your validation strategy carefully to ensure you can estimate your final model’s accuracy.
2. Explore different ways to measure the accuracy of the classifier. Examining the confusion matrix may be useful to see what it is doing, but you should also produce useful summaries, such as accuracy, precision, recall, F1 score, and so on. Choose summaries that would be the most useful to bean producers.
3. Which varieties are easiest to confuse with each other? Which are most obviously different? Perhaps you can look up the varieties to see if the results make sense; if your classifiers can’t distinguish between two bean varieties that appear very different, something might be wrong.

References

Koklu, M. and Ozkan, I.A., (2020), “Multiclass Classification of Dry Beans Using Computer Vision and Machine Learning Techniques.” Computers and Electronics in Agriculture, 174, 105507. https://doi.org/10.1016/j.compag.2020.105507