ISP Image Signal Processor in C

This repo is an implementation of ISP in c function, basically it is re-implemented from openISP in python. It's currently in very first version, I list some TODO list to make pipeline more robust.

And if you find something wrong in my implementation, welcome to let me know.

Function list

• Dead Pixel Correction

• Black Level Compensation

• Len Shading Correction

  (need calibration parameters)

• Anti-aliasing Noise Filter

• AWB gain Control

• CFA interpolation

  only implement rggb bayer pattern

• Color Correction

• Gamma Correction

  piecewise LUT

• Color Space Conversion

• Noise Filter for chroma

  • False Color Suppression

• Hue Saturation Control

• Noise Filter for Luma

  • Non local mean denoise
  • Bilateral Filter

• Edge Enhancement

• Contrast Brightness Control

Usage

build exe file,

mkdir build/
cd build 
cmake ..
cmake --build .

then isp_pipeline.exe will in prebuilt/ folder

test isp pipeline

isp_pipeline.exe test.raw

Configuration

The configuration can be a adjusted as input raw image.

In the config file , parameters are floats which are converted to fixed point type in C function to make operation in integer type.

config.json

{
    "width": 1920,
    "height": 1080,
    "image_bits": 10,
    "blc": {
        "bl_r": 0,
        "bl_gr": 0,
        "bl_gb": 0,
        "bl_b": 0,
        "alpha": 0.0,
        "beta": 0.0
    },
    "dpc": {
        "dead_thres": 30,
        "mode": "gradient"
    },
    "bayer_pattern": "rggb",
    "cfa_mode": "malvar",
    "awb_gain": {
        "r_gain": 1.5,
        "gr_gain": 1.0,
        "gb_gain": 1.0,
        "b_gain": 1.1
    },
    "ccm": [
        [
            1.0,
            0.0,
            0.0,
            0.0
        ],
        [
            0.0,
            1.0,
            0.0,
            0.0
        ],
        [
            0.0,
            0.0,
            1.0,
            0.0
        ]
    ],
    "gamma": 0.8,
    "csc": [
        [0.2568,0.5041,0.0979,16],
        [-0.1482,-0.291,0.4392,128],
        [0.4392,-0.3678,-0.0714,128]
    ],
    "nlm":{
        "Ds":4,
        "ds":1,
        "h":5
    },
    "bnf":{
        "dw":[
            [8,12,32,12,8],
            [12,64,128,64,12],
            [32,128,1024,128,32],
            [12,64,128,64,12],
            [8,12,32,12,8]
        ],
        "rw":[0,8,16,32],
        "rthres":[128,32,8]
    },
    "edge_filter":[
        [-1,0,-1,0,-1],
        [-1,0,8,0,-1],
        [-1,0,-1,0,-1]
    ],
    "eeh":{
        "gain_min":32,
        "gain_max":128,
        "thres_min":4,
        "thres_max":16,
        "em_clip_min":-64,
        "em_clip_max":64
    },
    "fcs":{
        "edge_min":32,
        "edge_max":64
    },
    "hsc":{
        "hue_offset":0,
        "saturation_gain":1.3 
    },
    "bcc":{
        "brightness":0,
        "contrast":1.1
    }


}

TODO

• board condition (I didn't handle any board condition in any function)

• lens shading correction (if I have parameters)

• CFA ( with more bayer pattern)

• color space conversion (with more yuv format)

• Non local mean denoise (optimization in exponential operation)

• bilateral filter (with non-fixed kernel and optimize in fixed point operation)

• False color suppression (the current result is kind of weird, need to find the reference algorithm)s

• Hue Saturation Control (the current result is kind of weird, need to find the reference algorithm)