This PR brings sinogram normalization (close #118 (closed)).
- Cuda implementation
- Unit tests
- Integrate in pipeline: config, process_config
- Test on datasets
- Tested on "ha_refill".
- In the pipeline (
FullFieldPipeline), normalization is done before building the sinogram. The reason is that in the case of half-tomo, the
build_sinostep modifies the sinogram structure, making normalization not work as expected. There is no really alternatives. The drawback is that any "normalization" that might need to happend after
build_sinowould need another intantiation/step.
- In the limited memory setting, normalization is done on host. It might be too slow for large chunks of data. Doing the processing on GPU might needs more work: either doing
VerticalShiftson GPU, or doing on normalization on GPU only if
- Only Chebyshev normalization is implemented. Splines normalization will have to wait.