| Popis: |
Meat Grading represents a major concern for the meat industry, as it sets the basis for pricing meat based on its quality. The current practice to assess marbling grade is done by human experts by looking at the ribeye muscle between the 12$\sp{th}$ and 13$\sp{th}$ rib bones, which is a very subjective approach and prone to high error rates. The main objective of the thesis is to grade the meat of live beef animals using ultrasound images of the ribeye muscle. Two classification algorithms are employed, namely, the Minimum Euclidean Distance classifiers and the k-means clustering algorithm. The input of the classifiers is the texture features vectors extracted from ultrasound images using the Co-occurrence matrix and the Laws masks. To improve the classification results, the texture features are transformed using the whitening transformation process to make them uncorrelated and to normalize their scale. A feature selection process is also used to select the best features that are more correlated to meat grade. To further improve the results and to reduce the uncertainty associated with the classification results, the classifiers are fused using the Generalized Bayesian Fusion approach. The GBF approach uses Bayes' theorem to combine the results of two or several classifiers. |
