Quality of beans is usually determined by visual inspection, which is subjective, laborious, and prone to error. Quality inspection is conducted after a bean classification process. Although manual classification is based on several features, automatic classification systems rely mainly on visual information. Thus, those systems are coincided to classify bean varieties which have different colours. In this paper, an automatic bean classification system is proposed using supervised learning algorithms, such as Support Vector Machine (SVM) and Random Forest (RF). Local and global features are evaluated to build classification models. Local features are calculated using the Scale-Invariant Transform Feature (SIFT), Local Binary Pattern (LBP) and Opponent-SIFT (OSIFT) descriptors, concatenated using three strategies (Early, Intermediate and Late fusion) and represented using the Bag-of-Feature (BoF) method. Codebooks with 500 and 1000 visual words are evaluated. Global features are defined based on the phenotype and morphology characteristics used by experts. A total of 600 bean images containing 39,040 individual bean seeds is used for evaluating the performance of classification models. Results showed that global features are more discriminant than local features because they are defined based on the application domain. The best accuracy of 98.5% using global features was obtained with RF, while the best performance using local features (accuracy of 95.2%) was obtained with SVM classifier, SIFT-LBP-OSIFT descriptors, intermediate fusion as concatenation strategy and a codebook with 1000 visual words. In addition, the proposed approach was compared against a strategy for bean classification of the-state-of-the-art which used Multilayer Perceptron neural network (MLP) and colour features. Experiments showed that the proposed local and global features achieved a better accuracy than the MLP based strategy which yielded an accuracy of 87.3%.