Analysis of the Recognition Effect on the Number of Spiral Arms in Spiral Galaxy Images Using ResNet

Abstract

The spiral arm information contained in spiral galaxy images, especially the number of spiral arms, is of great value for studying the structural evolution and dynamics of galaxies. Against the backdrop of explosive growth in galaxy observation data, how to quickly identify the number of spiral arms has become an important issue in the study of spiral galaxies. The research is based on the Galaxy Zoo DECaLS (Dark Energy Camera Legacy Survey) dataset and studies the ResNet (Residual Networks) model's method of identifying the number of spiral arms from spiral galaxy images. The experimental results show that the accuracy of the ResNet32 model is 83%, which is the best compared to network models such as ViT (Vision Transformer), EfficientNet, and DenseNet. In terms of recognition of different numbers of spiral arms, there is a strong relationship between recognition accuracy and the number of training samples. There are 6800 images with 2 spiral arms, with an F1-Score value of 0.9, while there are only 237 images with 4 spiral arms, with the lowest F1-Score value. The experiment further analyzed the recognition effect of fused traditional galaxy image features and found that the role of fused traditional galaxy image features in improving the recognition of spiral arms is limited.