残プロ 第-43回 ~カメラで撮ったレジュメをいい感じに~

台形検出+射影変換

レジュメなんかをデータ化する際，印刷機でスキャンできればいいのですが，忙しいとスマホのカメラで済ませることもあるかと思います．そうするとレジュメが曲がっていたり，机の余計な部分が移ってしまいます．
下記のプログラムでは，射影変換を行うことにより，スキャン風のいい感じな画像に補正してくれます．

import sys
import cv2
import numpy as np

def image2scan(src):
    # extract white
    hsv = cv2.cvtColor(src, cv2.COLOR_BGR2HSV)
    lower_white = np.array([0,0,100])
    upper_white = np.array([180,45,255])
    mask_white = cv2.inRange(hsv, lower_white, upper_white)
    res_white = cv2.bitwise_and(src, src, mask= mask_white)

    # detect trapezoid
    gray = cv2.cvtColor(res_white, cv2.COLOR_BGR2GRAY)
    tmp= cv2.findContours(gray, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    contours = tmp[0] if len(tmp) == 2 else tmp[1]
    areas = [cv2.contourArea(cnt) for cnt in contours]
    max_area = contours[np.argmax(areas)]
    epsilon = 0.1 * cv2.arcLength(max_area, True)
    approx = cv2.approxPolyDP(max_area, epsilon, True)

    # projection transform
    length = np.max(approx, axis=0)[0]
    cor = np.float32([[0, 0], [length[0], 0], [0, length[1]], length])
    approx_sum = [a[0][0]+a[0][1] for a in approx]
    sorted_sum = list(np.sort(approx_sum))
    sorted_idx = [approx_sum.index(i) for i in sorted_sum]
    pts = np.float32(np.array([
        approx[sorted_idx[0]][0], approx[sorted_idx[1]][0], 
        approx[sorted_idx[2]][0], approx[sorted_idx[3]][0]]))
    if pts[1][0] < pts[2][0]:
        pts = pts[[0, 2, 1, 3]]
    M = cv2.getPerspectiveTransform(pts, cor)
    dst = cv2.warpPerspective(src, M, (length[0], length[1]))

    return dst


def main():
    path = sys.argv[1]
    camera_image = cv2.imread(path)
    scan_like_image = image2scan(camera_image)
    try:
        write_path = sys.argv[2]
    except IndexError:
        write_path = "dst_" + path
    cv2.imwrite(write_path, scan_like_image)


if __name__ == '__main__':
    main()