Classificação de Personagens Animados usando Redes Neurais Convolucionais Profundas Pré-treinadas e Fine-tuning

Abstract

In this work, the author uses knowledge transfer and analyzes the use of the fine-tuning technique in convolutional neural networks for the character classification of the animated series “The Simpsons”. Knowledge transfer is carried out using deep neural networks already trained on large datasets of multiple classes, and the fine-tuning technique consists of carrying out a freezing of some weights and modifying others of an already trained model, on a new set of data. Studies about this technique report that some representations previously learned by a model are adapted to a new problem, which increases the performance of the model a little more. One concern present in this work was to avoid overfitting during the learning process of the models. For this, hyperparameters were introduced as regularization and dropout. Two neural network architectures were used, the use of fine-tuning in the models, on the data, caused a significant improvement in the results, improving the classification performance.