crop_marginalia.py

def white_band(img, side):
    """ 
    Determine best position to crop image

    Parameters:
    img (array): Binarised and skew corrected image
    type (str) : Side of image to crop (top, bottom, center, left or right)
    
    Returns:
    (int): Position on image to crop
    """
    h, w, c = img.shape
    pixel_size = 1 

    if side in ["top", "bottom"]:
        # Calculate white pixels at each slice   
        hist = np.sum(img/255, axis=1)[:, 0]   
        y = np.add.reduceat(hist, np.arange(0, len(hist), pixel_size))
        x = np.arange(0, len(hist),pixel_size)
        # Select indices to analyse image at specific area
        if side == "top":
            ind = x[x<h*0.2]    
        elif side == "bottom":
            ind = x[x>h*0.8]
    elif side == "center":
        # Calculate white pixels at each slice   
        hist = np.sum(img/255, axis=0)[:, 0]
        y = np.add.reduceat(hist, np.arange(0, len(hist), pixel_size))
        x = np.arange(0, len(hist), pixel_size)
        # Select indices to analyse image at specific area
        ind = x[np.logical_and(x<w*0.7, x>w*0.3)]
    elif side in ["right", "left"]:
        # Calculate white pixels at each slice   
        hist = np.sum(img/255, axis=0)[:, 0]   
        y = np.add.reduceat(hist, np.arange(0, len(hist), pixel_size))
        x = np.arange(0, len(hist),pixel_size)
        # Select indices to analyse image at specific area
        if side == "left":
            ind = x[x<w*0.2]    
        elif side == "right":
            ind = x[x>w*0.8]
    else:
        raise ValueError('Wrong side input')

    grd = np.gradient(y[ind])
    grd[np.logical_and(grd<np.std(grd)/2, grd>-np.std(grd)/2)] = 0  # Flatten noise within half sd
    nonzero = np.where(np.diff(grd)!=0)[0] # Determine largest white band
    crop = np.mean([nonzero[np.argmax(np.diff(nonzero))+1], 
                    nonzero[np.argmax(np.diff(nonzero))]]) # Center of white band 

    return round(crop+ind[0])