Selected problems of color image demosaicing in security imaging

The color TV security cameras are mostly based upon a single-sensor configuration with on-chip CFA (color filter array, Bayer configuration). The impact of different spatial sampling rasters in R G B channels causes an additional image distortion. In the field of security technology the identification task is among very frequent and above-mentioned sampling effect can affect it significantly. There are numerous interpolation algorithms suitable for the demosaicing - recalculation of full resolution R G B rasters. The paper is devoted to the comparison of several selected demosaicing algorithms usually applied in image enhancement of natural images where the subjective image quality is crucial. Our approach is oriented to the evaluation of identification limits in presence of noise and we are studying the noise resistivity of various demosaicing algorithms. We have chosen following demosaicing algorithms: adaptive homogeneity-directed demosaicing algorithm and demosaicing With directional filtering and a posteriori decision and three different types of noise: photon noise, thermal noise and salt pepper. In order to evaluate the subjective image quality two approaches are used. The first approach is a standard subjective quality assessment we have applied the identification threshold evaluation. As the second one the objective quality metric of noisy image subjective quality SSIM, Structural SIMilarity index is used. Both identification threshold and subjective quality dependencies are compared in order to evaluate a correlation degree. As test images we have chosen the car registration plate as an example of identification task esp. when the camera is recording images under low-light-level conditions. To keep the noise level precisely defined the test pictures were recorded almost noise-free and the noise was added consequently. The NEF (Nikon Electronic Format) picture format was applied to eliminate in-built image processing in camera.