Content-Based Galaxy Image Retrieval Using Convolutional Neural Networks

The morphology of galaxies can reflect their own physical information and evolutionary processes. However, with the continuous improvement and application of astronomical observation equipment, the number of galaxy images has also grown exponentially. Therefore, accurately retrieving specific images from a large dataset of galaxy images is a challenging task. Using content-based image retrieval and deep learning in galaxy images can easily retrieve similar images, aiding astronomers in their analysis. In this paper, we propose a method based on ResNet-50 for accurate galaxy image retrieval. First, we preprocessed data from the Galaxy Zoo project and categorized it into six classes. Then, we conducted a set of comparative experiments, training the ResNet-50 model using both cross-entropy loss function and triplet loss function. Experimental results show that the ResNet-50 model trained with the triplet loss function performed the best. The model achieved a mAP of 0.9366 and mAP@5 of 0.9538, which enable accurate retrieval of specific galaxy images.