Overview:
Image super resolution aims in recovering a high resolution image from a low resolution one.Our method directly learnsand end to endmapping between the high/low- resolution images. The mapping is represented as a deep Convolution Neural Network that takes the low resolution image as input and outputs a high-resolution one . We explore different network structures and parameter setting to achieve trade offs between speed and performance.
We conducted experiments from this paper.
We first tested different filters. We tried the following combinations:
NOTE: We have used T91 dataset for the following results.
A. 915 :
Comparing PSNRs between Bicubic and our SRCNN for YCrCCb:
Bicubic: 32.053059313091204
SRCNN: 33.56075266205568
Input:
Output:
B. 935 :
Comparing PSNRs between Bicubic and our SRCNN for YCrCCb:
Bicubic: 32.053059313091204
SRCNN: 33.68691645424939
Input:
Output:
C. 955 :
Comparing PSNRs between Bicubic and our SRCNN for YCrCCb:
Bicubic: 32.053059313091204
SRCNN: 33.616639192529114
Input:
Output:
Comparing three of them :