Uma proposta de arquitetura de rede neural convolucional intervalar para o processamento de imagens intervalares

Abstract

The general objective of this work is the proposal of an interval extension for convolutional neural networks and the analysis of their application in interval digital images in the context of pattern recognition in images, aiming for high accuracy and reliability in the results. In this work, we have an intervalal convolutional neural network with the purpose of controlling and automating the numerical error analysis, where the layers that compose the convolutional neural network are represented by equivalent operations through intervals. accuracy and classification. Traditional images are transformed into interval images, considering the neighborhood of 4 and 8 of their pixels. Then the network processing of these images is observed for the accuracy and error control. Then the concept of interval image slicing is inserted, looking for an improvement in the classification capacity of the network. Some cases and their effect on network accuracy are observed. Cross-Validation and Image Augmentation operations are still introduced to confirm overfiting and seek better network performance, respectively. It was observed that the allowable feature of slicing interval images was the best option for a better classification performance of the network in its current configurations.