Blind Image Super-resolution with Elaborate Degradation Modeling on Noise and Kernel

The Best of Both Worlds: a Framework for Combining Degradation Prediction with High Performance Super-Resolution Networks

“…B) t-SNE plots with each image coloured according to the compression applied. The colour intensity of each point corresponds to the compression level, which is in the range [30, 95] (higher is better) for JPEG and [20,40] (lower is better) for JM H.264. All three encoders are capable of separating compression magnitudes, but struggle to separate the two compression types when the compression is very low.…”

“…where θ D is the set of degradation parameters, which are unknown in practice. The function f can be expanded to consider the general set of degradations applied to I HR , yielding the 'classical' degradation model as follows [5,[17][18][19]21,22,[37][38][39][40][41][42]:…”

“…For a full exposition on the selection of each type of kernel, please refer to [15] [30,95] (following [15]). For JM H.264, images were compressed as single-frame YUV files where a random I-slice Quantization Parameter (QPI) were selected from the range [20,40], following [25]. All our models were trained on the LR images generated from the HR images of DIV2K [78] (800 images) and Flickr2K [79] (2,650 images).…”

“…Figure 6. t-SNE plots (perplexity value of40) showing the separation power of the different contrastive learning algorithms considered. Each dimension was independently normalised in the range [0, 1] after computing the t-SNE results.…”

“…Figure 10. t-SNE plots (perplexity value of40) showing the separation power of the different contrastive learning algorithms considered on the complex pipeline. All models were evaluated on the Iso/Aniso + Gaussian/Poisson + JPEG/JM testing scenario (4,944 images).…”