Percorrer por autor "Veiga, Bruno Tiago Silva"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Um classificador de divisões de interiores de casas para suportar agentes imobiliáriosPublication . Veiga, Bruno Tiago Silva; Pinto, Tiago Manuel Campelos FerreiraEveryday, Real Estate Agents retrieve pictures from houses that are then manually labelled and filtered to pick the best ones to present on their websites and better describe the homes they are selling. This is a time consuming task that delays the whole process of advertising homes and receiving propositions and could be solved with Computer Vision solutions. Machine Learning, more in particular Deep Learning, is a field that has been enjoying much success in the resolution of these types of problems, so it is a very promising solution to be used in this context. For that reason, this work proposes a neural network classifier for indoor images, that will save countless hours spent in these tasks. Moreover, to help achieve better results and provide additional contributions to the field, a study of the main architectures down the years, going from Convolutional Neural Networks to Transformers, and their performance for several datasets retrieved from online sources is also presented, settling all the information in one place and helping future researchers for similar tasks. The selected architectures for the case studies were the most relevant ones according to their popularity and performance, originating from different years, in order to provide an exploratory analysis of the evolution of the architectural aspects and the degree of impact these aspects have on the results obtained. These results confirmed that Transformers are indeed the ideal architecture for image classification tasks, especially in the real estate field, although Convolutional Neural Networks still play an important role, particularly in hybrid networks combining Transformers and Convolutional Neural Networks, where they are capable of achieving similar performance to purely Transformer networks. The best model achieved was able to predict images by training on a research dataset of 79500 images with an accuracy of 90%, which is quite good. This model was adapted and made available through an API. Some promising results were also obtained using a smaller and unbalanced private dataset of 10496 real estate images, where F1-scores of 87% and 86% were achieved for the multi-class and multi-label problems, respectively.