Objects detection in defective astronomical images using FPN-based CNN.
Models based on mask_rcnn_X_101_32x8d_FPN_3x by detectron2.
StarDet has been trained on defective astronomical images, allowing it to be used on images with background irregularities, star blooms and other defects. Example of detection on real images:
An example of detection on real astronomical images with the listed models is presented in example_real_images.ipynb. In model_training.ipynb you can see the training approach or train your own model with an example.
The model parameters and the CLAHE coefficients need to be adjusted for the images of the observatory of interest.
model_24e_anc4-64 params:
- Anchor_generator.sizes = [4, 8, 16, 32, 64]
- Num_epoch = 24
Evaluate and Losses of the model:
|
|
model_22e_anc8-128 params:
- Anchor_generator.sizes = [8, 16, 32, 64, 128]
- Num_epoch = 22
Evaluate and Losses of the model:
|
|