DATASET PIXEL METRICS: Inflexibility to dataset inputs with different numbers of channels.

[AlexSath]

Description

Currently, pixel statistics are calculated for training inputs to the de-noising model and ground truth de-noised images. These statistics allow for:

1. The normalization of inputs according to intensity percentiles calculated from a large sample will normalize any given input FLAME image (sample size n=1) to the pixel distribution from a large set of FLAME images.
2. Similarly, outputs from the de-noising model will be initialized to a pixel range from 0-1. While CARE-denoised FLAME images will never be fully quantitative, we can approximate quantitative pixel intensities by scaling the 0-1 range to intensity percentiles calculated from a large dataset of ground truth de-noised images (i.e. more frame accumulations).

To maximize the flexibility of the codebase for model training and evaluation, the ideal behavior would be for create_care_dataset.ipynb to aggregate pixel statistics in subsets of the larger dataset grouped by number of channels.

Example 1

Inputs to dataset: (N1,Y,X,3)

Current Behavior: Fields are saved to 'FLAME_Dataset' field in the dataset json.

• input_1pct, input_99pct, output_1pct, output_99pct: 1D array of size 3 (matching the number of channels in the input)

Example 2

Inputs to dataset: (N1,Y,X,3), (N2,Y,X,1), (N3,Y,X,16)

Current Behavior: Error; code crashes.

Desired Behavior: Each of the subfields saved to the 'FLAME_Dataset' field in the dataset json (input_1pct, input_99pct, output_1pct, output_99pct) should contain pixel statistics matching the number of channels in each input subgroup:

• 3chan: 1D array of size 3
• 1chan: 1D array of size 1 (or simply a float32 value)
• 16chan: 1D array of size 16

[AlexSath]

More information about how this is done for RGB images using mean RGB values for image normalization in Pytorch, please check out this answer on Stack Overflow.

The idea is to compute similar means and standard deviations for FLAME images so that one set of numbers can be used for all FLAME images as opposed to having to compute those numbers for every single dataset.