Simple segmentation

ml_simple_segmentation_scratch.py

import matplotlib.pyplot as plt
import cv2
import numpy as np

!wget  https://dl.dropbox.com/s/i1e1brycec3hy3k/pilot-view.jpeg -O pilot-view.png

# For this evalution, we assume that we load image with 3 channels.


def toRGB(img):
    return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)


def do_segmentation(input_img, threshold, max_value=255, min_value=0):
    N, M, C = input_img.shape
    image_out = np.zeros((N, M), dtype=np.uint8)
    label_out = np.zeros((N, M), dtype=np.uint8)

    for i in range(N):
        for j in range(M):
            if np.average(input_img[i, j, :]) >= threshold:
                image_out[i, j] = max_value
                label_out[i, j] = 1
            else:
                image_out[i, j] = min_value
                label_out[i, j] = 0

    return image_out, label_out


def extract_hist(img):
    list = []
    _, _, C = img.shape
    for i in range(C):
        list.append(cv2.calcHist([img], [
            i], None, [256], [0, 256]))
    return list, ["black"] if C == 1 else ["red", "green", "blue"]


pilot_view_img = toRGB(cv2.imread("pilot-view.png"))
print(pilot_view_img.shape)
M, N, C = pilot_view_img.shape

# For this evalution, we assume that we load image with 3 channels.
assert C == 3

# Show the interesting starts with contrast and high brightness background
pilot_view_neg_img = -1 * pilot_view_img + 255
plt.imshow(pilot_view_neg_img)
plt.show()

# The digital image has intensity valus of each pixel from 0~255, #size:256
intensity_values = np.arange(256)

# Extract the intensity distributions of all channels of each pixels
hist_list, colors = extract_hist(pilot_view_img)

# Extract pixels which have high intensity (brightness)
# The output image has pixel gray intensity 255 when the same pixel position of input has intensity >= 150, otherwise 0.
# For the label output the pixel given 255 has label "1", else "0".
result, label = do_segmentation(pilot_view_img, 150, 255, 0)
print(f"result shape {result.shape}")

plt.figure(figsize=(20, 20))
plt.subplot(131)
plt.title("origin")
plt.imshow(pilot_view_img)
plt.subplot(132)
plt.title("segmentation result")
plt.imshow(result)
# Bouns: Plot labels, the interesting pixel has been labeled as "1", otherwise be "0"
plt.subplot(133)
plt.title("outline by labels")
M, N = label.shape
label_image = np.zeros((M, N, C))
for i in range(M):
    for j in range(N):
        label_image[i, j, :] = 0 if label[i, j] == 0 else 255
plt.imshow(label_image)
plt.show()

# Let's compare the distribution of inensitiy distrubution

# Intensity distribution of origin image
plt.figure(figsize=(15, 4))
hist_orign_list, colors = extract_hist(pilot_view_img)
for i, hist, in enumerate(hist_orign_list):
    plt.subplot(1, len(hist_orign_list), i+1)
    plt.title(f"{colors[i]} channel #{i+1}")
    plt.bar(intensity_values, hist[:, 0], color=[colors[i]])
plt.show()

# Intensity distribution of result
plt.figure(figsize=(15, 4))
hist_result_list, colors = extract_hist(np.expand_dims(result, -1))
for i, hist, in enumerate(hist_result_list):
    plt.subplot(1, len(hist_result_list), i+1)
    plt.title(f"Channel {i+1}")
    plt.bar(intensity_values, hist[:, 0], color=[colors[i]])
plt.show()

# Plot continuous function for all channels intensity and greyscale of result
plt.figure(figsize=(15, 15))
all_list = hist_orign_list + hist_result_list
all_colors = ["r", "g", "b", "m"]
for i, hist in enumerate(all_list):
    plt.plot(intensity_values, hist, color=all_colors[i])

plt.axis([0, 300, 0, 1000])
plt.legend(["red ch", "green ch", "blue ch", "result graysale"])
plt.show()