Uso da inteligência artificial para a avaliação da cor instrumental de carne bovina

Abstract

Color plays an important role in the evaluation of meat quality, being the main sensory attribute evaluated by consumers and significantly influencing their purchasing decisions. The color of meat originates from the myoglobin pigment and the three chemical forms it can assume: deoxymyoglobin, oxymyoglobin, and metmyoglobin. Although colorimeters have a high cost and a limited representation of the meat surface, as well as limitations of proposed methods to quantify the redox forms of myoglobin, they are the scientifically accepted methods for instrumental color analysis. Alternatively, image processing using artificial intelligence is a possibility to overcome subjectivities and difficulties, in addition to being low cost, fast, objective, and without sample deterioration. Thus, in this dissertation, two articles were developed to explore, understand, and evaluate these phenomena. In the first article, color changes of the strip loin (Longissimus lumborum) were described during pigments' oxygenation and oxidation through traditional quantification methods (K/S and Krzywicki) of chemical forms (redox) and associated with colorimetric data. The ways to obtain 100% pigment standards for reference were also discussed. The redox forms of myoglobin were obtained according to the AMSA (2012) protocols. Oxygenation was evaluated for 60 minutes, refrigerated at 4ºC with readings every 10 minutes, and for discoloration during storage, monitored for 8 days (2ºC) with readings every 24 hours. It was concluded that color parameters and pigments are related but describe different information, such as color better describing the appearance of steaks and pigment changes, the temporal evolution. It was also concluded that the methods for predicting chemical forms differ, the isobestic points model (K/S) more coherently describing the behavior according to the expected chemical transformations, besides the c* and hº indices being good parameters to understand the changes in meat color. In the second article, the development of a computer vision system (CVS) using a Multi-Layer Perceptron (MLP) artificial neural network for quantifying myoglobin's chemical forms on the surface of strip loin steaks from images obtained from both a digital camera and a smartphone was evaluated, and the need for image calibration was tested. The same samples described in the first article were evaluated by both the colorimeter and the CVS. After algorithm training, the two best MLPs with an average accuracy of 92.62% ± 2.22 in classifying pigments were chosen. The MLPs' predictions were correlated with the mathematical Krzywicki method, demonstrating that CVSs were able to track pigment changes and, in some cases, had superior prediction results. In conclusion, this dissertation contributes to the literature that seeks to explain the instrumental color analysis of meat and how it relates to pigments in various real situations that beef can be subjected to, as well as the innovative approach of artificial intelligence for quantifying myoglobin's chemical forms, adapting methodologies to available technologies.