ISP learning (2)

One 、 Images bayer Format Introduction

bayer Format picture is Eastman · Kodak scientists Bryce Bayer Invented ,Bryce Bayer The Bayer array is widely used in digital image .
RGB Three primary colors are the basic elements of color pictures . The simplest acquisition method is to collect through the corresponding lens , The red filter passes through the red wavelength , The green filter goes through the green wavelength , Blue filter through blue wavelength . But the cost is high , It's not easy to make , Because all three filters must be aligned with each pixel .
bayer Format pictures with different colors set on a filter , By analyzing the perception of human eyes to color, we found that , The human eye is more sensitive to green , So general bayer The green format of the image is r and g The sum of pixels . That is similar to the present ISP How it works , By algorithm pair RGB The primary color of , Then get a better primary color map .

Pictured above ： Green is the sum of red and blue .
Bayer The format is the original picture inside the camera , The general suffix is .raw. Many software can view , such as PS. We took pictures with the camera and stored them on the memory card .jpeg Or pictures in other formats , from .raw Format conversion .

Bayer data , Its general format is ：
Odd scan line output GRGR……
Even scan line output BGBG……
According to the characteristics of large-area coloring with low response bandwidth of human eyes to color , There is no need to output each pixel at the same time 3 Color . therefore , When sampling data ,
The number of odd scan lines 1,2,3,4,… Pixels are sampled and output respectively G,R,G,R,… data ;
The... Of even scan lines 1,2,3,4,… Pixels are sampled and output respectively B,G,B,G,… data .
In actual processing , Of each pixel R,G,B The signal is composed of a certain color signal output by the pixel itself and other color signals output by adjacent pixels . This sampling method basically does not reduce the image quality at the same time , The sampling frequency can be reduced 60％ above .

Two 、bayer Format image sensor hardware

The structure of the image sensor is as follows , There is a metal isolation layer between each photosensitive pixel , The optical fiber passes through the microscope lens , After color filter filtering , Projected onto the photosensitive element of the corresponding porous silicon .

　 When Image Sensor When outputting data line by line , The sequence of pixels is GRGRGR…/BGBGBG…（ The order RGB）. Such an array Sensor Design , bring RGB The sensor is reduced to that of panchromatic sensor 1/3, As shown below .
Three primary colors , Green proportion 1/2 Blue and red are 1/4 1/4 The whole is that the collected three primary colors are original 1/3.
　

3、 ... and 、bayer Format interpolation algorithm implementation

3.1 Interpolation red blue Algorithm

At first , We collect RGB Data time , We use the same filter , Add different colors to the filter to collect . In order to get the RGB Elements , We need to fill in the missing by collecting the color 2 A color .
Because the human eye is most sensitive to green light , The reaction to purple light and red light is weak , So in order to achieve better picture quality , Need to be right G Special care .

　　　　(a) 　　　　　　　　　　　　 (b)
stay （a） And （b） in ,R and B Take the average value of the neighborhood .

　　　　 ( c) 　　　　　　　　　　　　 (d)
stay （c） And （d） in , Take... In the field 4 individual B or R The mean value of is used as the average value of the middle pixel B value .

3.1 Interpolation green Algorithm

　　　　 ( e) 　　　　　　　　　　　　 (f)
Above （c） And （d） in , The expansion is shown in the figure above （e） And （f） Intermediate pixel G The value of
（e） Intermediate pixel G The algorithm of value is as follows ：

（f） Intermediate pixel G The algorithm of value is as follows ：

I didn't understand G3+G4/2 This operation of . You need to check the information later .

The conversion of this part of the camera is used internally ADC perhaps ISP Accomplished , In order to reduce the cost, the manufacturer will inevitably make the image distorted . Of course, use an external processor to realize the conversion , If the processor is fast enough NB, Capable of pixel operation , Use the above algorithm to convert , All's well that ends well . However, the above algorithm will directly double the complexity of the algorithm , The speed will be limited . So for the Commission of speed , Can be directly through 4 field G Take the mean value to the of the middle pixel G value , Will double the rate , And the performance is very poor , Algorithm is as follows ：

Reference blog ：https://blog.csdn.net/qingzhuyuxian/article/details/82965054?ops_request_misc=&request_id=&biz_id=102&utm_term=%E5%9B%BE%E5%83%8Fbayer%E6%A0%BC%E5%BC%8F%E4%BB%8B%E7%BB%8D&utm_medium=distribute.pc_search_result.none-task-blog-2~all~sobaiduweb~default-1-82965054.nonecase&spm=1018.2226.3001.4187