Very Deep Super Resolution (VDSR)

Very Deep Super Resolution is a convolutional neural network that can perform single image super-resolution. It utilizes the differences in the high-frequency details between the low resolution and high-resolution images and estimates a residual image from it. The high-resolution image is taken as the reference image and the low-resolution image is obtained by upscaling with bicubic interpolation such that it matches the size of the reference image. The residual image contains the high-frequency details and is based only on the luminance information in this model.

The luminance channel of an image indicates the brightness level of a pixel as a linear combination of red, green and blue pixel values and VDSR uses this channel since the human perception can recognize a change in the brightness levels more effectively than the changes with the color which would depend upon the chrominance channels.

Fig 1: VDSR Generative process

[Image source: https://www.mathworks.com/help/images/single-image-super-resolution-using-deep-learning.html#mw_rtc_VeryDeepSuperResolutionDeepLearningExample_705D7EC3]

The VDSR network learns to predict the luminance of the residual image as the difference between the luminance of low-resolution and high-resolution images from the trained data.

From this, high-resolution images are obtained by combining the predicted residual image and the upscaled low-resolution image and converting it back to RGB colorspace.