Benchmarks for Debayer Algorithms on CPU and GPU

Here we present image demosaicing results for widely used debayer algorithms and compare CPU and GPU implementations.

Hardware

• Core i7-4770, 3.6 GHz, DDR-III 8 GB
• NVIDIA GeForce GTX 1080, driver 368.39
• Windows-7 (64-bit)

Software

• Fastvideo GPU Debayer (CPU and GPU)
  • HQLI (High Quality Linear Interpolation algorithm)
  • DFPD (Directional Filtering and a Posteriori Decision algorithm)
• Intel IPP-7.1 image processing library (CPU)
  • ippiCFAToRGB_8u_C1C3R (AN3 algorithm, multithreaded)
  • ippiDemosaicAHD_8u_C1C3R (AHD algorithm, multithreaded)
• OpenGL software on GeForce 9800 GX2 (according to McGuire article) for HQLI algorithm

We consider the following debayer algorithms: bilinear (AN3 from Intel IPP library), AHD (Adaptive Homogenuity Direction, from Intel IPP), HQLI (High Quality Linear Interpolation), DFPD (Directional Filtering and a Posteriori Decision).

Debayer Benchmarks on CPU and GPU

In that testing we measured computation time with and without host-to-device and device-to-host transfers to compare throughput for the above algorithms at CPU and GPU. Here we present averaged results for 8-bit images with 4K resolution (3840×2160). PSNR and SSIM results are calculated for standard Kodak data set.

Popular RAW converters consider AHD (AHDMF) algorithm to be industry standard. We have found how to get better results and not at the expense of performance. Here you can see moire patterns for the above algorithms.

You can also have a look at the latest Fastvideo SDK benchmarks for debayer algorithms HQLI, DFPD, MG on GPU. All tests are done on high performance mobile and desktop NVIDIA GPUs for 8-bit and 16-bit raw images.