Application and analysis of residual blocks in galaxy classification

Abstract

In astronomy, the automated galaxy classification method based on deep learning has significantly reduced the cost of manual annotation. The degradation problem in convolutional neural networks during galaxy classification tasks limits the accuracy improvement of deep models. Therefore, to address the issue of the model being too deep, which leads to a decrease in accuracy, the paper constructs the galaxy classification model using residual block structures. Specifically, this paper uses an improved ResNet as the backbone, stacking different numbers of residual blocks to extract input features. Meanwhile, smaller and deeper fully connected layers, regularization and activation functions, and Dropout layers are used to improve the model performance. For the best-performing ResNet152 model, the paper analyzes the classification report and confusion matrix and visualizes saliency maps and GradCAM heatmaps. Finally, the experimental results show that the introduction of residual blocks can increase the accuracy of the model by over 30%, and models with more residual blocks perform better, although the influence of the number of residual blocks on accuracy improvement is small. The visualization results show that the model can accurately segment the feature focus areas and points of interest in the original image. The model also focuses more on the central points with high planetary density by stacking multi-level residual blocks.