Classification of Bean (Phaseolus vulgaris L.) Landraces with Heterogeneous Seed Color using a Probabilistic Representation

Two of the most used techniques to characterize color in common bean landraces have been spectrophotometry and color analysis in digital images. The main limitation in previous works has mainly been that data have been obtained from specific points of homogeneous regions or mean of regions. A particular characteristic of native bean populations is that they comprise not only seeds of different colors but also of heterogeneous colors. We propose a computer vision system based on the use of histograms to represent the color properties from joint probability distributions of acquired color spaces that come from digital images in RGB and CIE 1976 L∗a∗b∗. We used 54 common bean landraces collected in different regions of the State of Oaxaca, Mexico. The classification accuracy of K-NN algorithm was 68.24%, 44.44%, and 53.80% with the spectrophotometer measures, RGB averages, and CIE 1976 L∗a∗b∗ averages respectively, while this same classifier achieved an average of 80% with histograms. Our results suggest that the two components regarding the chromaticity in CIE 1976 L∗a∗b∗ are enough to achieve the highest classification accuracy. Our proposal is not exclusive to classifying bean landraces; it might be used for fruit or vegetable color assessment.