Super-Resolution

In the following application examples, the input images were obtained by blurring and downsampling the original images. These downsampled images were then upsampled with various super-resolution methods.

General Images 2x Super-resolution

General Images 4x Super-resolution

Natural Images 2x Super-resolution

Natural Images 4x Super-resolution

Super-Resolution (Qualitative)

In the following application examples, the original images were used as input, so there exists no references against which to measure error.

General Images 4x Super-resolution

Super-Resolution Notes

The Gaussian kernel used to blur the images in the first group of examples had a standard deviation of 0.85 or 1.7 pixels and a support size of 1.5, 2.0, 3.5, or 4.0 pixels, depending on the downsampling scale and depending upon whether even- or odd-kernel downsampling was applied.

Bilinear, spline, bicubic, and Lanczos images were upsampled using ImageMagick version 6.6.

Perfect Resize 7.0 is a commercial image resizing application, previously known as Genuine Fractals. Note that this application does not align the pixels properly when it resizes an image, and therefore the reported mean squared errors in the Perfect Resize 7.0 examples are higher than they would be if the pixels were aligned.

The images in the second group of examples were obtained from The Weizmann Institute of Science Faculty of Mathematics and Computer Science Computer Vision Lab.

Denoising

In the following application examples, the input images were obtained by adding synthesized noise to the original images. These images were then denoised using various denoising methods.

The Recursive Conditional Means (RCM) method used in these examples is a method based upon Natural Scene Statistics.

General Images Denoising


Medium Noise Levels

High Noise Levels

Very High Noise Levels

Low Noise Levels

Natural Images Denoising


Medium Noise Levels

Very High Noise Levels

Low Noise Levels

Color channel prediction

In the following application examples, the image's red, green, or blue color channel was removed.

The missing color channel was then restored using one of two natural scene statistics methods:

  1. A nonparametric method, Recursive Conditional Means (RCM)
  2. A parametric method, Multiple Linear Regression (MLR)

On average across all images and channels the mean-squared error (MSE) for RCM is 10% lower than for MLR.