Galaxy morphology classification with densenet

Abstract

The classification of galaxies has always been an essential topic in astronomy, which can help to understand how galaxies form and evolve. This paper uses an effective deep-learning architecture, DenseNet-201, to classify galaxy morphology. Because galaxies are only concentrated in the center of the images, we preprocess the data in the way of reframing the images from the size of 256 × 256 × 3 to 224 × 224 × 3 which can eliminate all random noises like any other sub-object. The proposed method, DenseNet, connects all layers to each other. By using the DenseBlock+Transition structure, it realizes feature reuse and reduces the number of features, which could improve computational efficiency. We compare DenseNet-201 with VGG16 and MobileNetV2. VGG16 is very neat and contains multiple Conv->Conv->Max_pool structures, and the essential part of MobileNet is depthwise separable convolution. The whole network appears to be stacks of deeply detachable convolution modules. Applying DenseNet-201 to the dataset, Galaxy10 DECals, we achieved 84.1% accuracy of classification, higher than VGG16 of 79% and MobileNetV2 of 78%.