OpenCV OCR segmentation

0_segment.py

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


def main() -> None:
    f, (ax1, ax2) = plt.subplots(2, 1)

    im = cv2.imread('number_plate.png')

    # experiment with what works best here...
    # If the background can have bright colours but never white then np.min(im, axis=2) works well
    # If the text is a specific colour you could use HSV and do an inrange where the hue is correct and S/V match the text being foreground
    # If it's just a black/white then bgr2gray is simplest
    gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)

    # Find some threshold that works
    # OTSU is tempting, but can make the stroke vary depending on the background so if you can get a clean thresh with a value use that
    # invert as necessary
    # Add an empty border so our segmenting doesn't get unmatched start/ends
    _, thresh = cv2.threshold(gray, 100, 255, cv2.THRESH_BINARY | cv2.THRESH_BINARY_INV)
    thresh = cv2.copyMakeBorder(thresh, 1, 1, 1, 1, borderType=cv2.BORDER_CONSTANT, value=0)
    ax1.imshow(thresh)

    cols = cv2.reduce(thresh, 0, cv2.REDUCE_AVG)[0]
    ax1.plot(cols, color='r')

    # The peak threshold (height) is set somewhere that it will still detect the horizontal stroke of e.g. L, H, etc. as the same
    # letter but can filter noise. If the image is super noisy play with the morphology and/or blurring etc. of gray/thresh images
    gaps = cols < 10

    # edges now contains the start and end of each segment, with -1=segment start and +1=segment end
    edges = gaps[1:].astype(np.int) - gaps[:-1]
    ax1.plot(100 + edges * 50, color='g')  # transform signal to show on graph properly

    segment_starts = np.where(edges == -1)[0]
    segment_ends = np.where(edges == 1)[0]
    ax1.scatter(segment_starts, [40 for _ in segment_starts], color='blue')
    ax1.scatter(segment_ends, [40 for _ in segment_ends], color='orange')

    assert len(segment_starts) == len(segment_ends), 'Got unmatched segment start/ends. Do any characters/segmentations touch the edges of the image?'

    segments = []
    for start, end in zip(segment_starts, segment_ends):
        segments.append(im[:, max(0, start - 1):min(end + 1, im.shape[1] - 1)])

    ax2.imshow(np.hstack([
        cv2.copyMakeBorder(s, 0, 0, 0, 50 - s.shape[1], cv2.BORDER_CONSTANT, value=(255, 0, 255))
        for s in segments
    ]))

    plt.show()


if __name__ == '__main__':
    main()

Figure_1.png

number_plate.png