def findLightBlobs()

in ParametricLights.py [0:0]

• 45 lines of code
• 18 McCabe index (conditional complexity)

def findLightBlobs(I):
    # per light blob the ratio between max and lowest intensity
    energy_ratio = 0.3
    # (1-min_energy_wrt_max)*total_energy should be explained by the lights
    min_energy_wrt_max = 0.2

    min_pixels_left = np.shape(I)[0] * np.shape(I)[1] * 0.1
    if len(np.shape(I)) > 2 and np.shape(I)[2] > 1:
        rgb2gr = [0.2126, 0.7152, 0.0722]
        G=I[:,:,0]*0
        for c in range(3):
            G = I[:,:,c]*rgb2gr[c]
    else:
        G = I

    k = int(np.shape(I)[0] * 0.01 / 2) * 2 + 1
    G = cv2.GaussianBlur(G, (k, k), 0)
    W = panoweight(G)
    # 1 where pixels yet not yet accounted for
    mask = np.ones_like(W)
    all_lights = np.zeros_like(W)

    total_energy = np.sum(W)
    cnt = 1
    while 1:
        # find the next peak value
        seed = np.amax(W * mask)
        # find all values within energy_ratio of seed
        curr_light = np.zeros_like(W)
        curr_light[W * mask >= seed * energy_ratio] = 1

        # mask out the connected component to seed
        cc = measure.label(curr_light)
        # get the component where seed was
        seed_pos_1, seed_pos_2 = np.where(W * mask >= seed)
        # mark the light
        for c_ in cc[seed_pos_1, seed_pos_2]:
            if c_:
                all_lights[cc == c_] = cnt
                # mask it out for next round
                mask[cc == c_] = 0
        cnt += 1

        if np.sum(mask) < min_pixels_left or cnt > 15:
            break
        if np.sum(mask * W) < min_energy_wrt_max * total_energy:
            # there is less than min_energy_wrt_max left in the image
            break
    all_lights_merge = all_lights.copy()

    # Below are heuristics designed to tackle cases that occur in some images
    if 1:
        # Eliminate the cases where due to lack of sharp lights, almost the entire image/background is detected as one source.
        # This should be handled in the ambient component not light segments.
        props = measure.regionprops(all_lights_merge.astype(np.int))
        img_area = np.shape(I)[0] * np.shape(I)[1]
        for p in props:
            if p.area / img_area > 0.5 or p.major_axis_length / np.shape(I)[1] > 0.9:
                all_lights_merge[all_lights_merge == p.label] = 0
    if 1:
        # Merging of lights. Due to the thresholds, its possible to get nested lights in some cases.
        # Hence we merge such cases

        all_lights_merge[all_lights_merge > 0] = 1
        all_lights_merge = morphology.dilation(all_lights_merge, square(2)).astype(np.int)
        all_lights_merge = measure.label(all_lights_merge)

    return all_lights_merge, W