Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
169 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A Comparison of Deep Learning Architectures for Optical Galaxy Morphology Classification (2111.04353v1)

Published 8 Nov 2021 in cs.LG, astro-ph.GA, and astro-ph.IM

Abstract: The classification of galaxy morphology plays a crucial role in understanding galaxy formation and evolution. Traditionally, this process is done manually. The emergence of deep learning techniques has given room for the automation of this process. As such, this paper offers a comparison of deep learning architectures to determine which is best suited for optical galaxy morphology classification. Adapting the model training method proposed by Walmsley et al in 2021, the Zoobot Python library is used to train models to predict Galaxy Zoo DECaLS decision tree responses, made by volunteers, using EfficientNet B0, DenseNet121 and ResNet50 as core model architectures. The predicted results are then used to generate accuracy metrics per decision tree question to determine architecture performance. DenseNet121 was found to produce the best results, in terms of accuracy, with a reasonable training time. In future, further testing with more deep learning architectures could prove beneficial.

Citations (3)

Summary

  • The paper demonstrates that DenseNet121 achieves a balanced precision of 0.843 and recall of 0.851, outperforming EfficientNet B0 and ResNet50.
  • It employs tailored data augmentation and customized layers to effectively adapt deep learning architectures for galaxy morphology classification.
  • The study underscores the potential for automating astronomical classifications to efficiently manage large-scale Galaxy Zoo datasets.

A Comparative Study of Deep Learning Architectures for Optical Galaxy Morphology Classification

This paper presents a focused investigation into evaluating and comparing three well-regarded deep learning architectures—EfficientNet B0, DenseNet121, and ResNet50—in the context of optical galaxy morphology classification. As the field of astronomy is inundated with data due to the efficiency of modern astronomical surveys, the need for scalable and accurate automatic classification systems has become increasingly pressing. The Galaxy Zoo project, which crowd-sources morphological classifications from volunteer contributions, iterates this point through its voluminous dataset, hence forming an ideal testbed for deep learning models.

Core Methodology and Architectures

The research employs the Zoobot Python library, which follows a modified model training method based on the 2021 work of Walmsley et al. Within this framework, the authors utilized the Galaxy Zoo DECaLS data to adapt deep learning models for the specific task of predicting the decision tree-based classification responses, previously determined by volunteer assessments. EfficientNet B0, DenseNet121, and ResNet50 were employed as baselines allowing for a comparative analysis, particularly focusing on precision, recall, F1-score, and computational constraints such as training time.

The authors present a meticulous exploration where the generic model architecture integrates random rotations, flips, and cropping augmentations as input pre-processing steps, further engendering robustness in model predictions. To optimize morphological classification accuracy and computational efficiency, the significant architectural modifications include removing the top fully connected layer from each model and replacing it with layers specific to this task.

Results and Metrics

DenseNet121 emerges as the most proficient architecture amongst the three, offering a compelling balance between accuracy and computational efficiency. With weighted average scores of 0.843 for precision, 0.851 for recall, and 0.840 for F1-score, DenseNet121 outpaces EfficientNet B0 and ResNet50. It notably achieves this superior performance with an intermediate training time of 11.723 hours, further underpinning its efficacy. EfficientNet B0, while competitive in terms of accuracy, requires considerably longer training durations. ResNet50, though the fastest to train, falls short in precision, recall, and F1-score metrics.

Implications and Future Directions

From a practical standpoint, the paper indicates that DenseNet121 serves as a highly effective model for the automatic classification of galaxy morphology using optical images. This finding contributes positively to the ongoing quest for minimizing human labor in classification tasks and efficiently handling burgeoning datasets. Importantly, the research lays a solid groundwork for future explorations aimed at optimizing deep learning models for astronomy.

The implications are considerable, suggesting avenues for further exploration with additional architectures, potentially incorporating models like VGG19 or MobileNet. A broader architecture comparison could deliver more nuanced insights into the specific structural advantages of certain deep learning designs concerning astronomical tasks. Moreover, further studies might endeavor to apply these methodologies to other Galaxy Zoo projects, thus broadening the domain-specific applicability of these findings.

Conclusion

The paper adeptly addresses the pressing necessity for automated and scalable solutions in galaxy morphology classification by delivering a detailed, methodical comparison of deep learning architectures. DenseNet121 is identified as the prime candidate for future applications in this domain, thus representing a vital step forward in integrating AI techniques within astronomical research. While this paper sets a foundational understanding of architecture performance, continued experimentation and model refinement promise to push the boundaries of automatic classification capabilities further.

Github Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com