AAT JPEG Artefact Datection

Autodetects JPEG artefacts in categories of C100 (no artefacts), C95, C90, C85, C80, C70, C60, C40 and C20 (really terrible artefacts). The input is 1024x1024 large image. Image is converted to gray scale to omit redundant informations and reduce overall training and inference time. Model is self designed and doesn't use any other model as base.

Training stats

Around 60k of images was used during training.