Support for DLM (AIM) metric

[saswat0]

🚀 Feature

Support for Detail Losses and additive impairments as a differentiable loss function.

Motivation

VMAF is the de-facto benchmark for assessing quality in videos and is composed of two metrics: VIF (already implemented in torchmetrics) and ADM. To compute loss in videos using this metric, it'd be great to have its implementation. This is not only important for evaluating videos but also for individual images.

Pitch

Pixel-based difference measures such as PSNR and MSE correlate poorly with the human perception. To counter this, DLM[1] is a widely used for image quality assessment that doesn't depend on such pixel-wise statistics. Since all features of DLM are differentiable and can be ported to GPU instructions, it can serve as a great metric as a loss function.

1. Image Quality Assessment by Separately Evaluating Detail Losses and Additive Impairments

Alternatives

None

Additional context

NA

Related to #1245

[github-actions]

Hi! thanks for your contribution!, great first issue!