# libdebayer
A CUDA accelerated debayering library with C/C++ and Rust support.
## Dev Setup
```
nix develop
```
This enters a dev shell with all of the necessary build dependencies.
## Library Breakdown
### libdebayer
This is a simple C API that wraps the underlying CUDA kernels. The C
API assumes the images are already in GPU memory. libdebayer
implements 3 debayering algorithms:
- [Malvar 2004](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/Demosaicing_ICASSP04.pdf)
- Bilinear (same as builtin OpenCV)
- Saronic custom algorithm (TODO: more details here or delete since Malvar 2004 has the best performance)
### libdebayercpp
This provides a higher level C++ API that performs cudaMemcpy's from
host to device. An example of how to use this library is in
`benchmark/cpp`.
### libdebayer-rs
This provides a higher level Rust API with `TryFrom` traits to go from
an OpenCV `Mat` to a debayered `Mat`. An example of how to use this
library is in `rust/examples/test_benchmark.rs`. This test benchmark
is identical to the `benchmark/cpp` program.
### Run Rust Example
```
nix develop
cd rust
KODAK_FOLDER_PATH="<path-to-benchmark-kodak-files>" cargo run --example test_benchmark
```
## Benchmark
libdebayer is benchmarked against OpenCV and NPP (Nvidia Performance
Primitives). To run the benchmark for the Malvar 2004 implementation
run the following: `nix run .#kodak_benchmark_cpp`
### Benchmark Results
- OCV-EA gets Average PSNR: 28.616 dB
- NPP gets Average PSNR: 28.8522 dB
- libdebayer CUDA kernel gets Average PSNR: 33.4554 dB (+4.8394dB)
- Greatly reduces color fringing
#### Image Results
Example output from OpenCV-EA:
Example output from NPP:
Example output from our CUDA kernel:
