difference in calling dehazing using pip package

[bmabir17-asj]

In the readme it show the implementation of the dehazing as follows

$ import image_dehazer                                      # Load the library

$ HazeImg = cv2.imread('image_path', 0)                     # read input image
$ HazeCorrectedImg = image_dehazer.remove_haze(HazeImg) 

That means we are loading the image as grayscale and simply passing that image in remove_haze function.

But in example.py we are doing the following https://github.com/Utkarsh-Deshmukh/Single-Image-Dehazing-Python/blob/393a51439a56f90f78aba50d6ba5b8e7a9a184a6/src/example.py#L14-L40

That is we are using the color image and calculating Transmission Airlight and then passing those values to removeHaze(HazeImg, Transmission, A, 0.85)

Is the pip documentation wrong/outdated? or image_dehazer.remove_haze(HazeImg) function calculates them internally? If so why we are using grayscale image?

[Utkarsh-Deshmukh]

You are right. There is a mistake in the documentation. I have corrected the mistake. You should be using a color image, and not a grayscale image

[bmabir17-asj]

@Utkarsh-Deshmukh Okey, thanks. But what about this?

That is we are using the color image and calculating Transmission Airlight and then passing those values to removeHaze(HazeImg, Transmission, A, 0.85)

is this being calculated internally in pip package?

[bmabir17-asj]

On Further Inspection i found that it does calculate internally

def remove_haze(HazeImg, airlightEstimation_windowSze=15, boundaryConstraint_windowSze=3, C0=20, C1=300, regularize_lambda=0.1, sigma=0.5, delta=0.85):
    Dehazer = image_dehazer(airlightEstimation_windowSze=airlightEstimation_windowSze,
                            boundaryConstraint_windowSze=boundaryConstraint_windowSze, C0=C0, C1=C1,
                            regularize_lambda=regularize_lambda, sigma=sigma, delta=delta)
    HazeCorrectedImg = Dehazer.remove_haze(HazeImg)

Thanks for the awesome package :smile:

[yx-chan131]

Hi, why must it take color image but not a single channel image as input? @Utkarsh-Deshmukh

[Utkarsh-Deshmukh]

because in order to compute the airlight(which is later used to compute the haze transmission), we compute the dynamic range of the pixel within all the color channels