GPU-friendly Approximate Bilateral Filter for 3D Volume Data

This paper presents an approximate Bilateral Filter(BF) with a GPU-friendly architecture for 3D volume data. The bilateral filter (BF) for 3D volume data such as medical images highly costs due to an enormous number of voxels to be processed. Existing acceleration methods called constant-time, or O(1), BF are inappropriate for GPU processing because they consist of a combination of O(1) spatial filters not to fit to parallel processing. The proposed method realizes a fast approximation 3D-BF by focusing two points: (1) the BF is decomposed into multiple Gaussian Filters and (2) GPU processing is suitable for convolution. As a consequence, proposed method achieved fast and high approximate accuracy in various window size.