Currently, I am working on a super resolution project using a bi-cubic interpolation technique. I take a (512*512) as an original image for testing purpose then I reduced it (i.e. down-sampled) by 0.5 to be (256*256). Hence, the down-sampled image are used as an original image, and then it enlarged (i.e. up-sampled) by 2 to be (1024*1024). It is noticed that the output image became blurred and most details are disappeared. Due to this issue a bi-cubic interpolation technique is developed to enhance the output image, Meanwhile, it needs a sharpening filter to make edges sharp and illustrates the details very well...
two basic sharpening kernel are:
K1= [0 -1 0
-1 5 -1
0 -1 0]
K2= [-1 -1 -1
-1 9 -1
-1 -1 -1]
If you down-sample, of course you loose information. This information is lost and cannot be generated by any sophisticated interpolation technique.
Bi-cubic is a good one. I work with large 3D medical datasets, and for speed reasons, I use Catmull-Rom, which is performing slightly less than bi-cubic but is much faster.
How much information is lost also depends on the way you down-sample. It makes a difference whether you skip pixels, take the average or take the median for example.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing