A few days ago I've lanunched Optidash - ML-enhanced image optimization and processing API. Optidash builds on top of another product of mine (Pixaven) and runs entirely on Mac Pros (as pioneered by imgix).
While Optidash supports all major image formats for optimization and processing, I am mainly focused on JPEGs. All open-source JPEG optimizers share pretty much the same algo - create N copies of master image at different quality settings and, using various metrics (ssim/dssim/psnr), pick the variant with the "best" quality to size ratio.
Optidash takes a different approach. We use saliency detection to identify the most important area(s) of a master image. That basically tell us how the human eye would see the image and where it would look most likely. Once the saliency heatmap is computed, we crop that fragment and pass it to our Core ML model trained to predict optimal encoding settings. That approach also comes with a performance benefit - only the most salient areas are passed to the model (far less pixel data to process) and it also ensures we don't saturate pretty limited GPU memory we have available on Mac Pros (we use 2nd gen so D700, 6GB VRAM).
Estimating output Q value is one thing but we are also training additional models to help us determine optimal quantization table for a given salient region.
As I am still evaluating the above approach and general API design, I'd love to get some feedback.