Globular clusters (GC) are crucial for understanding galaxy formation and evolution. However, identifying them in large imagery datasets is a time-consuming task. This prompts the development of an automated GC detection algorithm. Although GC detection is fundamentally an object detection problem, the state-of-the-art object detection algorithms are unable to produce accurate results. Motivated by how GCs are identified by astronomers, we propose a deep neural network that fuses multiple views of raw imaging data and learns a better representation of the input image. The proposed network is then combined with YOLO object detection algorithm resulting in YOLO for Globular Cluster detection (YOLO-GC) model. Experimental results based on a real catalog of GCs in the M33 Galaxy showed that the proposed multi-view representation learning technique helps improve detection performance.

For details on data preprocessing, you can follow this link to execute: Data preprocessing

Clone the repository and install the required dependencies.

git clone https://github.com/administer03/YOLO-GC.git
cd YOLO-GC
pip install -r requirements.txt

Training script

python train.py \
    --img {image_size} \
    --epochs {num_epochs} \
    --batch-size {batch_size} \
    --data {dataset_config_path} \
    --weights {pretrained_weights} \
    --hyp {hyperparameter_file} \
    --cfg {model_config} \
    --device {device_id} \
    --exist-ok \
    --patience {early_stopping_patience} \
    --name {run_name} \
    --project {output_directory} \
    --fe_models {feature_extractor}

Training Parameters

Detection Script

python detect_modify.py \
    --imgsz {image_size} \
    --source {source_directory} \
    --conf-thres {confidence_threshold} \
    --iou-thres {iou_threshold} \
    --weights {model_weights} \
    --save-conf \
    --save-txt \
    --device {device_id} \
    --exist-ok \
    --raw-type {raw_type} \
    --name {run_name} \
    --project {output_directory} \
    --yoda_weights {yoda_weights_path} \
    --out_filter {output_filter}

Detection Parameters To perform inference with the pre-trined model, run the following command:

@inproceedings{
  singlor2023globular,
  title={Globular Cluster Detection in M33 Using Multiple Views Representation Learning},
  author={Singlor, Taned and Thawatdamrongkit, Phonphrm and Techa-Angkoon,
          Prapaporn and Suwannajak, Chutipong and Bootkrajang, Jakramate},
  booktitle={International Conference on Intelligent Data Engineering and Automated Learning},
  pages={323--331},
  year={2023},
  organization={Springer}
}