There are some relevant visibility improvement approaches presented in this post.
Beyond that, I quite like the breakdown of light travel and colour handling in the SeaThru paper, and the underlying ideas and data have since been expanded into SeaThru-NeRF and SeaSplat. To some extent they can be worked into real time processing if you have range estimates for different portions of your image (e.g. from stereo cameras), along with estimates of water colour and strength and colour of on-vehicle vs external sources of illumination, but they’re currently presented as post-processing approaches.
In terms of turbidity and haze there are various companies and researchers working on that, to varying degrees of success. There are some fundamental limitations to visibility range, but there is definitely scope to improve signal quality beyond a pure camera stream with temporal filtering (e.g. seeing which parts of the image persist over time), and (depending on what you’re trying to view) structure detection/enhancement (e.g. using edge detection algorithms). Traditional computer vision algorithms may have more deterministic/repeatable results on the same images, but it’s hard to avoid at least some amount of parameter tuning from data, at which point every approach becomes machine learning to some extent, and it’s mostly a question of whether there’s enough data and compute available to hand off more of the structure determination to the training process, or if some/much of that needs to be predefined in a hand-written algorithm. I’m aware that’s mostly theoretical insight - unfortunately I don’t have shareable practical examples available at front of mind.
For what it’s worth, the Blue Robotics Lumen lights are controlled with an exponential brightness curve so there’s extra fidelity in the low brightness regions, where there’s less backscatter and a lower risk of saturating the image sensor from bright scene regions. One way of dealing with “the light is shining back at me too much” is to just use less light (though of course that does reduce the total distance that can be seen, compared to more light in clear water).