Very long baseline interferometry (VLBI)} uses an array of physically disconnected telescopes to image astronomical objects. In VLBI imaging, a hidden astronomical image is recovered using measurements taken between pairs of telescopes, known as complex visibilities. A state-of-the-art approach for VLBI imaging is the regularized maximum likelihood method, which solves for an image that jointly maximizes the measured data log-likelihood and a hand-selected image regularizer.
We propose an alternative, data-driven approach that uses a convolutional neural network to reconstruct the hidden image from measurement data.
This work was presented as a poster and extended abstract at the WiCV Workshop at CVPR 2021.
To test our methods on simulated black hole images, you can download the dataset via Dropbox.
You can test our implementation of regularized maximum likelihood method using our python notebook demo.
First, download our pretrained model that was trained using the Fashion MNIST dataset with thermal noises added to the measurement data.
Next, you can test our neural network reconstruction network with complex visibilities using our python notebook demo.
You can train your own deep neural network for black hole imaging using our training script written in Python and TensorFlow. The model architechture is represented above.