Redes Neurais Convolucionais aplicadas a identificação e classificação de galáxias

Abstract

Surveys are automated telescopes strategically positioned on the Earth’s surface or in space, designed to scan the sky to build a detailed catalog of the celestial bodies observed in their field of view. Combined, these instruments generate terabytes of astronomical data, which makes it necessary to record and identify these objects due to the enormous volume of information produced. For this reason, qualified professionals and advanced deep learning computational techniques, such as Convolutional Neural Networks (CNNs), are needed to manage and process this enormous amount of data. This work investigates the applicability of several CNNs in classifying galaxies using images from the Galaxy10 SDSS Dataset, composed of 21,785 images, each with dimensions of 69x69 pixels. The tested frameworks, based on previous studies, were Gharat & Dandawate, EfficientNetV2-M, DenseNet121, ResNet50, AlexNet, VGG16 and VGG19. Preprocessing techniques such as data augmentation, resizing and contrast adjustment were applied in an active selected manner in different experiments, seeking to identify the most efficient combination in order to improve the results. The models were evaluated using analyses such as Accuracy, Loss, ROC-AUC and PR-AUC. The VGG19 architecture with application of the CutMix technique presented the best performance, achieving an accuracy of 0.90 and PR-AUC of 0.94. The results obtained not only confirmed the robustness of CNNs in the classification of astronomical data, but also highlighted the importance of choosing the model criteria and complementary techniques to optimize the classification process.