In the last years, innovative works, fueled by robust sky surveys and efficient methods of detection, rekindled the interest of the scientific community for Low Surface Brightness Galaxies, a peculiar class of galaxies with low superficial stellar density that are very faint and diffuse in optical images.
If the automatic detection of these galaxies in photometric surveys is already a complicated task, given their low surface brightness, searches for these objects suffer an aditional problem: the enormous amount of artifacts that are detected in the images and also have low surface brightness. With the growing amount of astrono- mical data, visual inspection to reject detected artifacts becomes impratical and it is necessary to develop efficient methods to separate Low Surface Brightness Galaxies from artifacts. Deep Learning methods, like Convolutional Neural Networks, are considered the state-of-the-art in several image classification problems.
In this work, we aim at comparing the performance of the Visual Tranformers network, which recently caused a huge impact in the literature for challenging the paradigm of the state-of-the-art, with Convolutional Neural Networks to identify Low Surface Brightness Galaxies. To do that, we implemented a Visual Transformers Network and compared it with the DeepShadows Network, a Convolutional Neural Network developed by Tanoglidis et al. (2021b). Both networks were trained using the only image set of these objects publicly avaliable, composed of 40000 images from the Dark Energy Survey. We verified that our model achieved slightly superior metrics in comparison with DeepShadows. Between those, when trained on the dataset using the PNG format, the standard model of our network achieved an accuracy = 0.929, which is 0.98% higher than the one of DeepShadows. However, when computing the uncertainties on the metrics of our method using the bootstrap method, we have noticied that the performance of our network was as good as the one obtained by DeepShadows when considering de 95% confidence interval of the metrics. Besides that, using images of the same dataset, but in the FITS format, we implemented a method to pre-processes the input data, which aims to improve the performance of the networks. This method consists of doing a contrast adjusment to highlight low surface brightness objects in the images. We noticed that the contrast adjustment contribuited both to the ViT standard model and DeepShadows to achieve a higher performance. However, the best training result with the FITS image set did not overcome the results of both networks when applied to the PNG dataset. Like Dosovitskiy et al. (2021) showed, with the increase of the datasets, the Visual Tranformers Networks can overcome the performance of Convolutional Neural Networks, therefore that seems to be an specially interesting paradigm to Vision Transformers.
Thesis (in protuguese) avaliable at https://drive.google.com/file/d/1-8ufORBAZ1wq8IUS_H35kHiGg_cutbaK/view?usp=sharing
Any doubts, please wrtite to manuel.stveras@gmail.com