Noise gating removes noise from image sequences using spatiotemporal coherence to identify statistically significant features of the images. In noise-limited image sequences (i.e. those with visible dark noise or photon shot noise) with sufficient spatial and temporal resolution to capture their subject, noise gating can improve SNR by roughly an order of magnitude (i.e. increase effective exposure time by 100x).
Image-noise-gate packages some helper routines and a single main entry point: noise_gate_batch, which accepts a collection of images as a 3-D NumPy array. The image batch is cut into smaller "cubies" that represent neighborhoods in the original data set. The cubies are apodized, Fourier transformed, and compared to a noise model (which may be constant noise, such as detector read noise; shot noise, for photon limited images; a hybrid model using shot noise and a constant noise floor; or multiplicative noise). The noise model is specific to the instrument used to collect the data; if you do not supply one, it is assembled from the data a posteriori. Fourier components that do not exceed the noise model by a specified factor are zeroed ("gate" method) or smoothly attenuated with a Wiener filter ("wiener" method). The cleaned data are then inverse Fourier transformed and finally reassembled into the original shape.
The method leaves an apodized border 2/3 the size of the "cubies", along each edge of the data -- so if you specify a neighborhood size of 12x12x12, a swath 8 pixels wide is invalidated at each border of the data set.
The basic operation is to noise-gate a batch of images that can all fit in RAM. That operation is implemented with noise_gate_batch(). Large-batch and stream processing are also possible and are implemented with wrapper functions around noise_gate_batch.
noise_gate_batch(data_cube, cubesize=12, model='hybrid', factor=2.0)
Not implemented yet
Not implemented yet
The code works on two key principles: (1) the Fourier transform concentrates recognizable image features, but not noise -- easing discrimination between noise and features, without direct reference to the position of a feature within a neighborhood; (2) in image sequences that are strongly affected by noise, it is possible to estimate both the overall noise level and how it varies across the dataset, from the dataset itself.
The noise model generation involves treating the Fourier spectra of all the "cubies" in the data set, as a statistical ensemble. From this ensemble, the code generates a noise spectrum specific to the data source, using the particular model specified ( constant, shot, hybrid, or multiplicative ).
The noise gating part involves estimating a location-specific noise model using the contents of each cubie and the supplied noise spectrum. Finally, Fourier components of each cubie that do not exceed the corresponding noise model are either zeroed or attenuated.
For more details, see DeForest 2017 (ApJ 838, 155).
This package was ported and adapted from a collection of Perl/PDL routines "noise_gate") first written in 2017.
Commercial use of the method is subject to U.S. Patent No. 10181181.
This version of the method makes use of NumPy operations (such as fft) in the per-neighborhood hot spot, so it is ~2x slower than the PDL code. Dropping into C and using fftw and other per-pixel primitive operations might bring it up to parity with PDL.
Craig DeForest Southwest Research Institute [email protected]