Sistema de visão computacional para avaliação física de cafés (Coffea arabica L.) de diferentes colorações

Abstract

The color of coffee varies due to species, storage conditions and type of processing. The evaluation of coffee bean color is done by visual inspection by trained panelists, a very subjective method. Thus, trading companies demand for rapid and objective methods for color evaluation. The computer vision system emerges as an alternative for verifying the color of coffee beans, therefore, this work aims at building a computer vision system for identifying the different colors in coffee beans. To do so, we performed a conversion of the RGB of digital cameras, given that they are capable of obtaining information in pixels in the color parameters L * a * b * for each pix el of the digital image, thus obtaining a more complete set of information on the coffee bean color. In order to create the computer vision system, we used: a dark metal box, digital camera, lighting system and an image processing software based on neural networks. For the construction of the transformation model, we used color cards and, for the pattern recognition, we acquired coffee samples in different colors: off-white, sugarcane green, green and blue-green. Each color class presented 30 samples containing 50g each. In addition, we used a classification system (Bayesian classifier) to separate the samples into classes and verify the efficiency of the created system. The transformation model stood out with an error of only 1.20 + 1.24 for training and of 1.15 + 1.1 for testing. The Bayesian classifying system was efficient for classifying the samples used for validation within the classes. The samples were classified within the color classes, which represents an efficiency of 100%. With the results obtained, we verified that the off-white samples showed a high value of the L*, a* and b* parameters, which represents an approximation to white, while the sugarcane green and green samples showed intermediate parameter values; however the b* parameter for sugar cane green was higher, showing a certain yellowing of the sample. The green samples presented lower a* values, demonstrating that it is closer to the green color. However, the blue-green samples showed low values of L*, a* e b*, which represents the approximation to green and blue colors. The different color samples were efficiently classified, demonstrating the efficiency of the computer vision system. The system implemented in this work can be expanded to cooperatives and companies in the near future, providing a faster and more objective way to evaluate color.