Uso de redes neurais convolucionais para identificação de manifestações patológicas no concreto armado

Abstract

The early identification of pathological manifestations in reinforced concrete structures is essential for the safety and durability of buildings. However, the conventional inspection process, based on visual and tactile analysis, can be time-consuming. Therefore, this study aims to investigate the application of Convolutional Neural Networks (CNNs) for the automated classification of pathological manifestations in reinforced concrete structures. The data was collected from various sources: the International Congress on Pathologies and Rehabilitation of Constructions (CINPAR), technical reports of building inspections, and photographic records from the author. Six pre-trained CNNs were tested using Transfer Learning, namely: VGG16, VGG19, ResNet50, ResNet50V2, EfficientNetV2B0, and EfficientNetV2B3, analyzing their efficiency in feature extraction and subsequent anomaly classification. The results showed that the combination of the ResNet50 extractor with CNNs achieved the best performance, with an accuracy of 97.89%, sensitivity of 97.15%, and a Matthews Correlation Coefficient (MCC) of 94.86%. The proposed approach demonstrated high efficiency in identifying structural defects such as cracks, corrosion, and efflorescence, contributing to faster and more reliable inspections of reinforced concrete structures. Future work suggests expanding the database and refining image preprocessing techniques.