Aprendizado profundo aplicado à classificação de peças de xadrez.

Abstract

This research presents a solution based on deep learning neural networks for the classification of chess pieces. The goal is to evaluate the applicability of these algorithms in contexts such as robotics, for example. With that in mind, different neural network models were trained, with different architectures and hyperparameters. Later, the metrics accuracy, precision, recall and f1-score were calculated for each trained model and these metrics were compared to define the model with the best performance. The goal of each model was to correctly classify the input image into one of thirteen classes, these classes being 12 chess pieces and one class representing the empty space of the board. In this way, a model with 99.15% accuracy in the recognition of chess pieces was reached, this model was based on the MobileNet architecture and had as the best parameters found: learning rate of 0.00019, 256 neurons in the dense layer and 43 of the first layers frozen. Furthermore, pre-trained weights from ImageNet were used in this model. These results show the effectiveness of deep neural networks in classifying chess piece images.