search

csjunxu / TWSC-ECCV2018

Matlab Code for "A Trilateral Weighted Sparse Coding Scheme for Real-World Image Denoising, ECCV 2018".
Other
91 stars 26 forks source link

About Par.lambda1 #2

Open ghost opened 5 years ago

ghost commented 5 years ago

Hi, Xu. I'm now working on sparse coding image denoising, however, I find that in your program, Par.lambda1 is always set to 0, which means you actually don't use ADMM to find the sulotion. So the function you define in the papar has the close form solution, and it can be simply solved by a soft-thresholding operation. Therefore, the weights, W1 and W2, do not actually affect the solution, and W2 only benefits the aggregation process. I'm not sure whether my above understanding is correct or not. Thank you.

csjunxu commented 5 years ago

W2 is the key in this model. I just set the lambda1=0 for simpler model, in which W1 is an identity matrix. The experiments show that W1 is not that important for real world noise, but is essential to deal with images with clear noise variances in different channels.

I tested the case when lambda1 is not equal to 0, but forgot the results. I will check again this case in future.

Hope my answer is meaningful.

Thanks.