New Image > Lighting > Dehaze tool: ready for primetime!

Details in the comments!  Let me copy+paste key details:

Automatic image dehazing is an area of active study.  It is of special interest to automative manufacturers, who need access to fast, high-quality dehazing algorithms for automative safety system cameras.

PD's implementation is currently based on a 2018 paper in the open-access journal "Mathematical Problems in Engineering" by Zhong Luan, Hao Zeng, Yuanyuan Shang, Zhuhong Shao, and Hui Ding.  The paper is titled "Fast Video Dehazing Using Per-Pixel Minimum Adjustment" and you can you download it here (link good as of September 2021): https://www.hindawi.com/journals/mpe/2018/9241629/

Many thanks to Zhong Luan et al for sharing their work under a Creative Commons Attribution License.

I read quite a few dehazing papers from the past two decades before settling on this one for PhotoDemon's implementation.  One of the biggest issues I have in PD development is implementing methods that meet a few strict criteria:
 1) no massive 3rd-party libraries (e.g. OpenCV is out) or amd64-only libraries
 2) good run-time performance without modern intrinsics, including SIMD (necessary for XP compatibility)
 3) better performance than O(n^2) because considering (1) and (2), I still have to make the function fast enough to be useable on 20+ megapixel photos, even on PCs with strong RAM/CPU limitations.

These criteria get harder to meet every year, especially as I attempt to add things to PD beyond the usual low-hanging photo editing fruit of brightness, contrast, etc.

Anyway, I liked this paper because most dehazing approaches rely on either massive regional analysis on each pixel (which is incredibly slow), optimizing lengthy differential equations on-the-fly (which is even worse), or machine-learning algorithms that require huge training sets (ugh no).  For example, this link describes a classical dehaze implementation based on the original dark-channel prior theory by Kaiming He: https://sites.google.com/site/computervisionadinastoica/final-project

As the authors at that link state: "Running the algorithm on any larger images would require days of computation... He, Sun, and Tang mentioned using the Preconditioned Conjugate Gradient algorithm as a solver, allowing them to process a 600x400 pixel image in roughly 10-20 seconds."

10-20 seconds for a 600x400 image!  Not gonna work for PhotoDemon.

Convesely, the paper by Zhong Luan et al uses a much-simpler per-pixel analysis with good - and predictable! - results, excellent performance, low memory requirements, and many areas of potential further optimization that play nicely with VB6's limitations.  Using this strategy, a 10-megapixel image processes in less than a second, regardless of inputs.  Much better.

As such, while this work is heavily inspired by the original paper, it is *not* identical.  (Nothing built in VB ever is lol.)  I have heavily commented the code to note deliberate modifications and/or enhancements, and I am open to input on further changes.  As with other functions in PD, I generally try to err on the side of "more subtle, useful real-world results" vs "intense HDR-style changes" that look impressive in a paper but produce wildly unrealistic-looking results on things like regular iPhone photos.

Ideas for further improvements are of course welcome.

Author: tannerhelland