|
|
|
|
|
|
by iamaaditya
3464 days ago
|
|
|
If it would be sufficient to have high PSNR only in the 'regions of interest', then you could use something like this [1], which uses a CNN model to predict a map, and a multiple JPEG encoding to achieve variable quality compression. [1] https://github.com/iamaaditya/image-compression-cnn |
|
|
MozJPEG is actually using something similar called trellis quantization or soft-thresholding. The idea is to further quantize DCT coefficients in noisy areas. There are some limitations [1] though. But it's not an issue with locally sensitive metrics and can be improved with edge detection.
Another challenge here is to estimate the current image quality and the target error.
[1] https://github.com/mozilla/mozjpeg/issues/129