COSMICA: A Novel Dataset for Astronomical Object Detection with Evaluation Across Diverse Detection Architectures

Area of Science:

• Astronomy and Astrophysics
• Computer Science
• Artificial Intelligence

Background:

• Accurate celestial object detection in astronomical images is crucial but challenging due to noise and varying object characteristics.
• Traditional methods often fall short in real-time applications and complex observational conditions.

Purpose of the Study:

• To introduce COSMICA, a curated dataset of manually annotated astronomical images from amateur observations.
• To evaluate the performance of modern object detection models for real-time astronomical object identification.

Main Methods:

• Investigated YOLOv8, YOLOv9, YOLOv11, EfficientDet-Lite0, and MobileNetV3-FasterRCNN-FPN on the COSMICA dataset.
• Evaluated models using metrics such as mAP, precision, recall, and inference speed under consistent experimental conditions.

Main Results:

• YOLOv11 achieved the highest overall accuracy and computational efficiency among the tested models.
• All evaluated models showed varying degrees of success in detecting comets, galaxies, nebulae, and globular clusters.

Conclusions:

• Deep learning models, particularly YOLOv11, show significant promise for real-time celestial object detection in observational astronomy.
• Domain-specific datasets like COSMICA are vital for training robust AI models for astronomical applications.