Noise reduction in HDR merging

[demoulinv]

Description

Update the HDR merging algorithm to reduce the resulting noise when the majority of the original LDR images are under exposed and very dark.
Optionally compute and output 3 light mask (darklights, highlights, no midlights) for the purpose of the LDR stack of views analysis.

linked to alicevision/Meshroom#2072

Features list

• Update the interface of the selectTargetView function by returning the index of the selected reference view.
• Add 2 new parameters (min and max significant values) in the main LDR_to_HDR_merge executable.
• Update hdr process method with the new algorithm.
• Output 3 light masks (darklights, highlights, no midlights).

Implementation remarks

The main idea of the algorithm update is to determine for each channel of each pixel an optimal set of views to be merged. Assuming that the views are sorted from the shortest to the longest exposure (from the darkest to the brightest views) and a reference view has been selected, starting from the index of the reference view the index to start is decremented while the channel value is higher than the minimum significant value. Starting from the computed first index incremented by 1, the last index is incremented while the channel value is lower than the maximum significant value.