eason-sun/color-demosaicing

Color-Demosaicing

Problem Definition:

Most digital color cameras today are equipped with a single sensor array (CMOS or CCD) whose surface is covered by a color flter array (CFA). Each unit of CFA contains a spectrally selective flter so that it samples only one of the three primary colors (red, green and blue).

Original Image:

Task 1: Implemented a simulated Bayer CFA. Use it to sample color images.

Result:

Task 2: Reconstruct the image based on CFA data using bilinear interpolation.

Result:

Observation:

The perceptual quality is acceptable. If we look close enough, several artifacts are observed:

1. The main body of CN Tower has several color fringes.
2. False color is created on the surface of the buildings.
3. The output image is blurry due to loss of sharpness.

Potential Cause:

This simple interpolation algorithm performs well on surface with homogeneous colors. On area with high spatial frequency where two colors separated by a sharp transition (details of edges & borders), false colors are generated. Basically, the bilinear interpolation is operating spatial averaging like a 2-dimensional low pass flter. Therefore, it will also result in blurry image.

Task 3: The Adaptive Color Plane Interpolation (Hamilton and Adams.) method is chosen to overcome the problems addressed in bilinear interpolation.

Result:

Pros:

1. The right image is apparently much sharper and contrasty and contains more edge details.
2. From the main body of the CN Tower, we can clearly see that the color fringes are absent.
3. If we focus on the high frequency area (the surface of buildings), we could observed that the false color generated by bilinear interpolation is largely reduced.

Cons:

1. There are noticeably many red noise pixels on the black areas and yellow & green noise pixels on the white areas. (They can be easily removed by noise reduction tools)
2. If we look at the small building on the right, we could capture absurd repetitive patterns in transition areas (zipper efect).

Conclusion:

From the comparison above. We could conclude that the Hamilton and Adams method for interpolation outperforms its bilinear counterpart both perceptually and statistically. In principle, Adaptive Color Plane Interpolation uses classifers to estimate the gradient in horizontal, vertical and diagonal directions such that the high spacial frequency information is better realized upon the preferred orientation. However zipper efect and noise pixels are noticeable, which leads to the future improvements.