qw245 / blindfuse Goto Github PK

When I was reproducing the algorithm of your paper,” Blind Model-Based Fusion of Multi-band and Panchromatic Images”, I found that pansharpening by a factor of 4 is having a better result than pansharpening by a factor of 2, which is very counterintuitive.

I first guessed this is due to the bad choice of lambda in the task of pansharpening by a factor of 2. However, when I finetuned lambda, i.e. the constant 1e-3, to another value in Line 32:
invCov = 1e-3*eye(L); in SylvesterFusion.m, the average PSNR did not change too much.

Therefore, I was wondering how to exactly finetune the parameters for the task of pansharpening by a factor of 2 so that its performance can be reasonably higher than the performance from pansharpening by a factor of 4.

If you want to reproduce the counterintuitive results, please run Pavia_small_2x.m and Pavia_small_4x.m in the dropbox link: https://www.dropbox.com/sh/zjkecgkryoyoklh/AACHWuptGCrpjYufqO7DStwea?dl=0. You will find the first experiment generates the average PSNR: 31.0754 dB and the second experiment generates the average PSNR: 31.7673 dB.