APRENDIZADO PROFUNDO PARA CLASSIFICAÇÃO DE IMAGENS: EXPLORANDO O USO CONJUNTO DE REDES NEURAIS PARA CLASSIFICAÇÃO E REMOÇÃO DE RUÍDO EM IMAGENS DE DÍGITOS MANUSCRITOS
Artificial Intelligence is the Computer Science field that seeks to understand and develop intelligent machines, capable of taking efficient actions to solve problems in unknown scenarios. Among its subfields is Deep Learning, represented by biological brain inspired algorithms for artificial neural networks development. Its relevance lies in the possibility of application to different problems in different areas. As an example, and the research focus, image classification is cited — characterized by the need to assign one (or more) label(s) to an image based on the information it represents. However, it is observed that neural networks are negatively affected by the noise presence in data. This impact can be reduced by applying pre-processing methods for noise removal. Among the existing techniques in the literature are the autoencoder neural networks. In this scenario, two neural networks are used together to solve the noisy image classification problem. Considering this, the research questions the use of information from the classification model to develop and adapt the noise removal model. Therefore, the general objective is to explore the possibility of an image classifier accuracy increase using an autoencoder noise remover, adapting it by introducing the classification error metric into its learning algorithm. The research scope is limited to the handwritten digits image classification problem — chosen for its relevance and tangible study complexity. The method used consists of experimental research consisting of four procedures, namely: (I) analyze the performance of a neural network in solving the digit classification problem in noise-free images; (II) evaluate the classifier performance when used on noisy images; (III) study the impact on classification performance caused by image pre-processing; (IV) explore the noise removal model adaptation and analyze its impact on the studied classification problem. Among the results, it is highlighted that the adaptation explored can improve the correct classifications rate for pre-processed noisy images. Furthermore, there is evidence that adaptation makes the development of noise removal models more efficient. For future work is mentioned an explanatory research development aiming to formally detail the results obtained. Another possibility is to replicate the adaptation explored in other scopes that make joint use of two (or more) neural networks.
Keywords: Artificial Intelligence; Image Classification; Noise Removal.
O presente repositório é reservado ao armazenamento dos recursos utilizados durante o Trabalho de Conclusão de Curso (TCC) do Bacharelado em Ciência da Computação ofertado pelo IFC - Campus Blumenau. Os documentos como a monografia (pré e pós-correção) e a apresentação do TCC são disponibilizados no diretório "Documentos", acompanhados pelos respectivos modelos LaTex. Os códigos para a pesquisa experimental estão disponibilizados no diretório "Pesquisa Experimental" e para executá-los, verifique se todas as bibliotecas necessárias estão instaladas.
Destaque: esta pesquisa de TCC originou o artigo entitulado "Classificação de Dígitos Manuscritos em Imagens Ruidosas: explorando o uso e a adaptação conjunta de redes neurais para classificação e remoção de ruído". O trabalho será apresentado em sessão técnica no Computer on the Beach 2024 em 10/04/2024.