[PYTHON] [Machine learning] I tried to do something like passing an image

Hello. This is Yuki Ando (@ holiday1173).

In this article, using a type of clustering method called the K-means method introduced in the previous article, [pathization (vectorization)](https: // school) of images, which is often used in Illustrator, etc. .dhw.co.jp/word/design/vector.html) I tried something like that.

Overview

Here's what I did in the last article:

1. Clustering the pixel value of each pixel of the image using K-means

2. Assign the average pixel value in the cluster for each cluster, reduce the number of colors and draw

Previous article → [Machine learning] Explaining color quantization

Here is the result of the actual processing. (Left: Before processing, Right: After processing)

The image on the right shows the result of K-means processing with 9 clusters. Since the number of clusters = the number of colors, the processed image is drawn in only 9 colors. The image looks like a slightly flat coloring book.

In the previous article, we used only pixel values as input for K-means, so if the colors were similar, pixels in a slightly distant area of the image would be divided into the same cluster.

So, I thought that if I clustered the pixel coordinate information in addition to the pixel value, I could get a path-like vector-like image, so I tried it.

result

Here is the code I used.

import os
import argparse
import numpy as np
import cv2


if __name__ == "__main__":
    #Specify the path as an argument
    parser = argparse.ArgumentParser()
    parser.add_argument('--img', '-i', help='The path of the image you want to process')
    parser.add_argument('--output', '-o', help='Directory to output the result image')
    args = parser.parse_args()

    #Loading images
    img = cv2.imread(args.img)

    #Image as it is k-Since it cannot be applied to means, shape(Number of pixels, 3(BGR))Conversion to
    Z1 = img.reshape((-1, 3))
    
    """
Get xy coordinates for each pixel
    """
    #Vertical and horizontal width of the image
    h, w = img.shape[:-1]
    #Two-dimensional list of xy coordinates of images
    Z2 = [[i, j] for i in range(h) for j in range(w)]
    
    # RGB +Concatenate xy coordinates
    Z = np.hstack((Z1, Z2))
    # np.Convert to float32 type
    Z = np.float32(Z)

    # k-means termination condition
    #Use default value
    criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 10, 1.0)
    #Number of groups after split
    for K in range(5, 21, 3):
        # K = 10
        # k-means processing
        _, label, center = cv2.kmeans(
            Z, K, None, criteria, 10, cv2.KMEANS_RANDOM_CENTERS)

        # np.Convert to uint8 type
        center = np.uint8(center)
        #Assign intra-group mean for each group
        res = center[label.flatten()]
        res = res[:, :-2]
        #Restore the original image size
        res2 = res.reshape((img.shape))

        output_path = os.path.join(args.output, 'Parrots_K{:02}.png'.format(K))
        #Save image
        cv2.imwrite(output_path, res2)

Here are the main processes added to the previous code.

"""
Get xy coordinates for each pixel
"""
#Vertical and horizontal width of the image
h, w = img.shape[:-1]
#Two-dimensional list of xy coordinates of images
Z2 = [[i, j] for i in range(h) for j in range(w)]

This part of the process gets the xy coordinates for each pixel. By combining the xy coordinates obtained here with the list of pixel values, a list that concatenates RGB + xy coordinates is generated. First, get the vertical width (h) and horizontal width (w) of img. The xy coordinate list is created with the line Z2 = ~ as [[0, 0], [0, 1], [0, 2] ...].

# RGB +Concatenate xy coordinates
Z = np.hstack((Z1, Z2))

This line combines the list of pixel values and xy coordinates, and the data to be input to K-means is completed.

Below are the results. The original image

K = 5

K = 8

K = 11

K = 14

K = 17

K = 20

Not only the pixel value but also the position is taken into consideration when dividing into clusters, so it looks like a path!

reference

- Image </ b>
The image data used this time can be downloaded from the link below.
http://www.ess.ic.kanagawa-it.ac.jp/app_images_j.html