Marcos Tidball
Low Surface Brightness Galaxies (LSBGs) are believed to be a significant portion of the matter in the local Universe, with many questions about them still left unanswered. A big problem when it comes to studying these objects is how challenging it is to observe them, not only for how diffuse they are, but also due to the large amount of data that has to be analysed. Thus, this work focuses on developing a method that uses Convolutional Neural Networks (CNNs) in order to automate the detection of these galaxies. We make use of DeepScan, an astronomical source extraction package designed to detect low surface brightness features in astronomical data, with fine-tuned thresholds to crop stamps out of images. Since neural networks require a large number of training data to make precise predictions and there is a small number of available real LSBGs, we opted to use simulations of this kind of galaxy. Both the DeepScan stamps of non-LSBGs and simulated LSBGs are then fed to a pre-trained CNN in order for it to be trained in a classification problem regarding the detection of LSBGs and non-LSBGs. The project is still ongoing and we are currently experimenting with different pre-trained networks in order to achieve the best results.
The basis of a CNN is a convolution operation. It is based on the usage of a filter, which is a grid of values. This filter is passed through the image, multiplying and adding together the values of the filter with the values of the area of the image inside the filter. Through this we obtain a convolved image, which lets us detect different characteristics on an image. What is special about a CNN is its ability to learn the best filter values according to its training data. This is possible through a process called “training”, where the network is fed with labelled images and has to learn how to classify them. A loss function calculates how wrong the network is and through a backpropagation algorithm it is possible to adjust the filters in order to minimize the loss function.
We created a pipeline to detect LSBGs automatically. The pipeline is called “Deepfuse” and is divided into two parts: the first uses DeepScan, a source extractor tool, to detect low surface brightness features in astronomical images. The second part then implements a CNN that classifies these stamps. DeepScan is a library made for detecting low surface brightness features and works by gradually grouping together neighbouring pixels based on their signal-to-noise ratio. It was necessary to apply some thresholds in surface brightness and effective radius in order to specialize the algorithm to detect the objects that interest us.
In order to train the CNN we used two classes of images: positives (LSBGs) and negatives (non-LSBGs). Due to the small amount of real LSBGs available we used simulations for the positive images. For negative images, on the other hand, we visually inspected DeepScan detections and selected the ones that are not LSBGs. The simulated galaxies were inserted in the same astronomical images from where we extracted the non-LSBGs. This way it is possible to have the same image properties for both samples. Since we have only a small amount of positive and negative images, we made use of transfer learning: using a CNN that was pre-trained on large amounts of data and retraining just part of it in new data. This makes it possible to maintain the filters that detect general features while also specializing more easily on new data.
The project is still ongoing. Currently we are performing tests with different networks and parameters in order to find the combination with the best results.