Fig. 5
From: ConIQA: A deep learning method for perceptual image quality assessment with limited data
Image transformations used for consistency training in ConIQA. Two types of transformations were used to augment (a) samples from our HQA1k dataset. We considered (b) horizontal flipping, and (c) a novel augmentation method based on weighted averaging. The latter transformation generates a new sample from an existing pair, wherein the target is fixed and the rendering is replaced by a weighted average of the target and the rendering. The weight, \(\lambda\), is a small random number drawn from a uniform distribution in the [0, 0.05] range.
