Estudo de viabilidade da classificação visual de danos aparentes em postes de concreto utilizando visão computacional

Abstract

Traditionally, the structural condition of electrical distribution assets is assessed through field inspections and visual analyses conducted by technical teams, a process that is time-consuming and presents limitations for the overall evaluation of the electrical network. In this context, this work proposes the use of image recognition techniques associated with a trained computational model, specifically MobileNetV2, to classify images of concrete utility poles as being in good or poor condition, thereby supporting maintenance assessment. The project was implemented using the Python programming language, open-source libraries for image processing, and a computational model previously trained on the ImageNet dataset. Transfer learning was applied while keeping the layers responsible for feature extraction frozen. A dataset assembled by the author containing 150 images, equally distributed between the two classes, was used. The images were resized, normalized, and subjected to data augmentation techniques. During the evaluation of the test set, the model trained with cropped images achieved accuracies of 95% and 85% for the first and second image distributions, respectively. However, errors still occurred in the identification of poles in poor condition. Based on the results, it was concluded that the proposed approach has potential as an auxiliary inspection tool; however, it is not yet sufficiently reliable for autonomous classification. Therefore, expanding the image dataset and improving data preprocessing procedures are still necessary.