Implementação em hardware de rede neural para reconhecimento e classificação de imagens

Abstract

The goal of this project is the development of a digital circuit in a FPGA board capable of implementing functionally a neural network for image recognition and categorization. This neural network application shall be given by performing the classification of clothing articles images in existing databases. Its purpose is to apply the advantages of parallel processing, flexibility and performance of FPGAs to generate an effective solution with low computational requirements in the creation of artificial intelligence applications, striving to reduce financial and energy costs for its usage. To this end, a software neural network will be created utilizing the programming language Python, where all the relevant fundamental equations and linear algebra concepts will be implemented without using any existing machine learning libraries (such as Keras, TensorFlow or PyTorch) since it will be used as reference for the digital circuit design, which will then be implemented in a FPGA development kit. All relevant data will be loaded in the FPGA memory and processed internally só that the results containing the image information and neural network predictions are presented to the end user.