Fast 3D Non Local Means Filtering using Convolutions CPU/GPU

Domain Decomposed CPU/GPU implementation of Non Local Means Filtering in 3D using Convolution

A fast 3D non local means filter cast using the convn function for improved scalability.

Also features a bounding box serialiser to fit volumes into memory.

Usage:

[nlmf] = serialGPUNLMF(domain,locSize,DoS,k,npasses,gpuFlag)

nlmf: the output domain

domain: the input 3D array

locSize: the subdomain size in X,Y,Z. e.g. locsize=256 will partition the domain into subvolumes of 256^

DoS: the degree of smoothing (commonly 0.05 to 0.2)

k: the filter window size as an odd number (3 is minimum, performance penality high for large numbers, recommend 5-7)

npasses: the number of times the filter is applied

gpuFlag: boolean, use gpu or not. (convn is commonly known to be slower on a GPU for cases where locSize is small. At this point, while GPU arrays work with this code, it is seemingly not very fast).

[nlmf] = serialGPUNLMF(domain,256,0.1,5,2,0)