Extract a slide image from a photo

extract_slide.py

# -*- coding: utf-8 -*-

# ============================================================ #
# 写真からスライド領域を切り出して保存。透視変換もおこなう
#
# Usage: python extract_slide.py YOUR_PHOTO
#
# ============================================================ #

import cv2
import numpy as np

# ============================================================ #
# 画像の準備
# ============================================================ #

import sys
img_path = sys.argv[1]

# 画像のロード
img = cv2.imread(img_path)
# BGR から RGB に変換
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# ============================================================ #
# 画像の前処理
# ============================================================ #

# グレースケール化 + ガウスぼかし
img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img_preprocessed = cv2.GaussianBlur(img_gray, (5, 5), 0)

# 大津の方式で二値化
_, img_binary = cv2.threshold(img_preprocessed, 0, 255,
                              cv2.THRESH_BINARY + cv2.THRESH_OTSU)

# ============================================================ #
# 輪郭の抽出
# http://opencv.jp/opencv-2.1/cpp/structural_analysis_and_shape_descriptors.html
# http://docs.opencv.org/3.1.0/dd/d49/tutorial_py_contour_features.html
# ============================================================ #

contours = cv2.findContours(np.copy(img_binary),
                            cv2.RETR_EXTERNAL,
                            cv2.CHAIN_APPROX_SIMPLE)[1]

# ============================================================ #
# 輪郭のなかからスライドっぽいものだけを抜き出す
# ============================================================ #

def simplify_contour(contour):
    """
    Simplify the given contour using Douglas-Peucker algorithm
    """
    hull = cv2.convexHull(contour)
    epsilon = 0.1 * cv2.arcLength(hull, True)
    return cv2.approxPolyDP(hull, epsilon, True)

contours_simplified = [simplify_contour(contour) for contour in contours]

# 四点からなるモノだけ抽出
contours_rectangles = filter(lambda contour: len(contour) == 4, contours_simplified)

if contours_rectangles:
    # 最もスライドらしい領域をひとつだけ選ぶ
    def slide_likelihood(contour):
        """
        Compute the score how the given contour looks like a slide
        """
        # 一番大きな輪郭がスライドという仮定の下で、一番面積の大きな輪郭を算出
        # （グリーンの定理で面積を求めているらしい）
        return cv2.contourArea(contour)

    largest_contour_index = np.argmax([slide_likelihood(contour)
                                       for contour in contours_rectangles])

    # ============================================================ #
    # 画像を切り取る
    # ============================================================ #

    # 最もスライドらしい領域の bounding box を算出
    largest_contour = contours_rectangles[largest_contour_index]
    x, y, w, h = cv2.boundingRect(largest_contour)

    def normalize_contour(contour):
        """
        まず x の値で並び替え、その次に y の値で並び替える
        """
        left_part = contour[np.argsort(contour[:, 0])][0:2]
        right_part = contour[np.argsort(contour[:, 0])][2:4]

        left_top, left_bottom = left_part[np.argsort(left_part[:, 1])]
        right_top, right_bottom = right_part[np.argsort(right_part[:, 1])]

        return np.array([left_top, left_bottom, right_top, right_bottom])

    # 透視変換用に座標情報を用意
    slide_rect = np.array([[0, 0],
                           [0, h],
                           [w, 0],
                           [w, h]])
    from_rect = normalize_contour(largest_contour.reshape(4, 2)).astype(np.float32)
    to_rect = normalize_contour(slide_rect).astype(np.float32)

    # 透視変換をかける
    M = cv2.getPerspectiveTransform(from_rect, to_rect)
    warped_img = cv2.warpPerspective(img, M, (w, h))

    # 切り取ってみる
    cropped_img = img[y:y+h, x:x+w]

    # 切り取った/透視変換した画像を保存する
    cv2.imwrite(img_path + ".cropped.jpg",
                cv2.cvtColor(cropped_img, cv2.COLOR_RGB2BGR))
    cv2.imwrite(img_path + ".warped.jpg",
                cv2.cvtColor(warped_img, cv2.COLOR_RGB2BGR))

    print("Finished processing" + img_path + ". Saved the cropped/warped image")
else:
    print("Slide not found in " + img_path)