Recognition of coffee roasting degree using a computer vision system

Abstract

The definition of the coffee roasting degree is mainly based on the coloring of beans and is directly related to the beverage quality. This bean color reading usually occurs by visual inspection process or by using traditional instruments with scope limitations. Thus, the aim of this study was to construct a computational vision model that compares color patterns in CIE L*a*b* and grayscale with the numerical scale of roasting defined by Specialty Coffee Association of America. Artificial neural networks were used as a color transformation model and quadratic/cubic polynomial regression models and neural models for roasting index approximation. For whole beans, the applied Tukey test (95% of confidence level) showed that the neural model outperformed the polynomial ones for roasting index approximation, getting a R2 factor of 0.99. For ground beans, the quadratic polynomial grayscale model was the best predictor, showing an average error of 0.93. Therefore, the proposed system is considered as effective in the identification and approximation of coffee bean color allowing greater automation and reliability in roasting degree analysis.