For the convolutional neural network SRCNN for super-resolution reconstruction, the only preprocessing step is bicubic (bicubic interpolation), which constructs the given low-resolution image to the target pixel size.
For example, if we have a 32*32 pixel picture and want to reconstruct it into a 64*64 pixel resolution picture, we first use the bicubic interpolation method to expand the 32*32 pixel picture to 64*64, and then the resulting image Put into SRCNN for next steps.
The understanding of interpolation method is as follows:
For example, a 2*2 pixel image is doubled to 4*4
1. Insert pixels
2. Fill the inserted pixel with brightness value, taking dest(I, J) as an example, the specific steps are:
(1) According to the mapping relationship, find the closest pixel point ori(u,v) of dest(I,J) in the original image origin, where ori(u,v)=90
(2) Take ori(u,v) as the center and radiate 16 points outward. The coordinates of these 16 points are ori(i+u,j+v), and the value range of i,j is -1<= i<=2;-1<=j<=2, if the original image does not have these pixels, it can be filled and completed.
(3) Operate these 16 points according to a certain relationship (similar to convolution) to obtain the brightness value at dest(I,J).