ISAR Image Registration using a CNN-Bi-LSTM Architecture in PyTorch

PyTorch implementation of this paper

Notes/Todo

• Currently processing 120x120 images. Original image sizes are around 250x170

Installation/Running

Create conda environment:

conda env create -f environment.yml

Running gradio dashboard using pre-trained weights:

python dashboard.py

Training

Extract data and place into a directory called data/. Modify utils/constants.py for a different location. The data has already been labeled in labels.csv.

To train, run:

python train.py --train --test <directory_to_store_weights>

Modify the weights directory in dashboard.py for visualization.

Hyperparams and Performance

200 epochs, learning rate 0.0001, batch size 128:

• Sequence length 3: .7306
• Sequence length 5: Test acc .7753
• Sequence length 10: Test acc .7593

Data Specifications

GEOEYE: Images #1-17630
--Range Resolution (y-axis): 0.0047 m
--Cross-Range Resolution (x-axis): 0.0047 m
SPASE: Images #17631-32840
-Images #17631-21509,#25083-32840
--Range Resolution (y-axis): 0.0062 m
--Cross-Range Resolution (x-axis): 0.0062 m
-Images #21510-25082
--Range Resolution (y-axis): 0.0057 m
--Cross-Range Resolution (x-axis): 0.0058 m

Images are shown in log-scale and are thresholded below at 60 dB below the peak.