minAreaRect OpenCV [Python] 返回的裁剪矩形

Question

OpenCV 中的minAreaRect返回一個旋轉的矩形。 如何裁剪矩形內的這部分圖像？

boxPoints返回旋轉矩形角點的坐標，因此可以通過遍歷框內的點來訪問像素，但是在 Python 中有更快的裁剪方法嗎？

編輯

請參閱下面我的答案中的code 。

Answer 1

這是一個執行此任務的函數：

import cv2
import numpy as np

def crop_minAreaRect(img, rect):

    # rotate img
    angle = rect[2]
    rows,cols = img.shape[0], img.shape[1]
    M = cv2.getRotationMatrix2D((cols/2,rows/2),angle,1)
    img_rot = cv2.warpAffine(img,M,(cols,rows))

    # rotate bounding box
    rect0 = (rect[0], rect[1], 0.0) 
    box = cv2.boxPoints(rect0)
    pts = np.int0(cv2.transform(np.array([box]), M))[0]    
    pts[pts < 0] = 0

    # crop
    img_crop = img_rot[pts[1][1]:pts[0][1], 
                       pts[1][0]:pts[2][0]]

    return img_crop

這里是一個示例用法

# generate image
img = np.zeros((1000, 1000), dtype=np.uint8)
img = cv2.line(img,(400,400),(511,511),1,120)
img = cv2.line(img,(300,300),(700,500),1,120)

# find contours / rectangle
_,contours,_ = cv2.findContours(img, 1, 1)
rect = cv2.minAreaRect(contours[0])

# crop
img_croped = crop_minAreaRect(img, rect)

# show
import matplotlib.pylab as plt
plt.figure()
plt.subplot(1,2,1)
plt.imshow(img)
plt.subplot(1,2,2)
plt.imshow(img_croped)
plt.show()

這是輸出

Answer 2

這是執行上述任務的代碼。 為了加快這個過程，不是首先旋轉整個圖像並裁剪，而是首先裁剪具有旋轉矩形的部分圖像，然后旋轉，然后再次裁剪以給出最終結果。

# Let cnt be the contour and img be the input

rect = cv2.minAreaRect(cnt)  
box = cv2.boxPoints(rect) 
box = np.int0(box)

W = rect[1][0]
H = rect[1][1]

Xs = [i[0] for i in box]
Ys = [i[1] for i in box]
x1 = min(Xs)
x2 = max(Xs)
y1 = min(Ys)
y2 = max(Ys)

angle = rect[2]
if angle < -45:
    angle += 90

# Center of rectangle in source image
center = ((x1+x2)/2,(y1+y2)/2)
# Size of the upright rectangle bounding the rotated rectangle
size = (x2-x1, y2-y1)
M = cv2.getRotationMatrix2D((size[0]/2, size[1]/2), angle, 1.0)
# Cropped upright rectangle
cropped = cv2.getRectSubPix(img, size, center)
cropped = cv2.warpAffine(cropped, M, size)
croppedW = H if H > W else W
croppedH = H if H < W else W
# Final cropped & rotated rectangle
croppedRotated = cv2.getRectSubPix(cropped, (int(croppedW),int(croppedH)), (size[0]/2, size[1]/2))

Answer 3

@AbdulFatir 提出了一個很好的解決方案，但正如評論所述（@Randika @epinal），它對我來說也不是很有效，所以我稍微修改了它，它似乎對我的情況有用。 這是我正在使用的圖像。

im, contours, hierarchy = cv2.findContours(open_mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
print("num of contours: {}".format(len(contours)))


mult = 1.2   # I wanted to show an area slightly larger than my min rectangle set this to one if you don't
img_box = cv2.cvtColor(img.copy(), cv2.COLOR_GRAY2BGR)
for cnt in contours:
    rect = cv2.minAreaRect(cnt)
    box = cv2.boxPoints(rect)
    box = np.int0(box)
    cv2.drawContours(img_box, [box], 0, (0,255,0), 2) # this was mostly for debugging you may omit

    W = rect[1][0]
    H = rect[1][1]

    Xs = [i[0] for i in box]
    Ys = [i[1] for i in box]
    x1 = min(Xs)
    x2 = max(Xs)
    y1 = min(Ys)
    y2 = max(Ys)

    rotated = False
    angle = rect[2]

    if angle < -45:
        angle+=90
        rotated = True

    center = (int((x1+x2)/2), int((y1+y2)/2))
    size = (int(mult*(x2-x1)),int(mult*(y2-y1)))
    cv2.circle(img_box, center, 10, (0,255,0), -1) #again this was mostly for debugging purposes

    M = cv2.getRotationMatrix2D((size[0]/2, size[1]/2), angle, 1.0)

    cropped = cv2.getRectSubPix(img_box, size, center)    
    cropped = cv2.warpAffine(cropped, M, size)

    croppedW = W if not rotated else H 
    croppedH = H if not rotated else W

    croppedRotated = cv2.getRectSubPix(cropped, (int(croppedW*mult), int(croppedH*mult)), (size[0]/2, size[1]/2))

    plt.imshow(croppedRotated)
    plt.show()

plt.imshow(img_box)
plt.show()

這應該會產生一系列像這樣的圖像：

它還會給出這樣的結果圖像：

Answer 4

你沒有給出示例代碼，所以我也沒有代碼回答。 您可以進行如下操作：

1. 從矩形的角，確定相對於水平軸的旋轉角度 alpha。
2. 通過 alpha 旋轉圖像，使裁剪的矩形與圖像邊框平行。 確保臨時圖像的尺寸更大，以免丟失信息（參見： Rotate image without cropping OpenCV ）
3. 使用 numpy 切片裁剪圖像（參見： 如何使用 Python 在 OpenCV 中裁剪圖像）
4. 將圖像旋轉回 -alpha。

Answer 5

不幸的是， Oliver Wilken的回答並沒有導致顯示的圖像。 也許是因為不同的 openCV 版本？ 這是我采用的版本，它增加了幾個功能：

• 矩形的縮放和填充，即獲得原始矩形之外的部分
• 結果圖像的角度可以相對於矩形進行配置，即 0 或 90 [度] 的角度將水平或垂直返回矩形
• 返回平移矩陣以旋轉其他事物，例如點、線、...
• 用於 numpy 和 openCV 數組索引和矩形操作的輔助函數

代碼

import cv2
import numpy as np


def img_rectangle_cut(img, rect=None, angle=None):
    """Translate an image, defined by a rectangle. The image is cropped to the size of the rectangle
    and the cropped image can be rotated.
    The rect must be of the from (tuple(center_xy), tuple(width_xy), angle).
    The angle are in degrees.
    PARAMETER
    ---------
    img: ndarray
    rect: tuple, optional
        define the region of interest. If None, it takes the whole picture
    angle: float, optional
        angle of the output image in respect to the rectangle.
        I.e. angle=0 will return an image where the rectangle is parallel to the image array axes
        If None, no rotation is applied.
    RETURNS
    -------
    img_return: ndarray
    rect_return: tuple
        the rectangle in the returned image
    t_matrix: ndarray
        the translation matrix
    """
    if rect is None:
        if angle is None:
            angle = 0
        rect = (tuple(np.array(img.shape) * .5), img.shape, 0)
    box = cv2.boxPoints(rect)

    rect_target = rect_rotate(rect, angle=angle)
    pts_target = cv2.boxPoints(rect_target)

    # get max dimensions
    size_target = np.int0(np.ceil(np.max(pts_target, axis=0) - np.min(pts_target, axis=0)))

    # translation matrix
    t_matrix = cv2.getAffineTransform(box[:3].astype(np.float32),
                                      pts_target[:3].astype(np.float32))

    # cv2 needs the image transposed
    img_target = cv2.warpAffine(cv2.transpose(img), t_matrix, tuple(size_target))

    # undo transpose
    img_target = cv2.transpose(img_target)
    return img_target, rect_target, t_matrix


def reshape_cv(x, axis=-1):
    """openCV and numpy have a different array indexing (row, cols) vs (cols, rows), compensate it here."""
    if axis < 0:
        axis = len(x.shape) + axis
    return np.array(x).astype(np.float32)[(*[slice(None)] * axis, slice(None, None, -1))]

def connect(x):
    """Connect data for a polar or closed loop plot, i.e. np.append(x, [x[0]], axis=0)."""
    if isinstance(x, np.ma.MaskedArray):
        return np.ma.append(x, [x[0]], axis=0)
    else:
        return np.append(x, [x[0]], axis=0)


def transform_np(x, t_matrix):
    """Apply a transform on a openCV indexed array and return a numpy indexed array."""
    return transform_cv2np(reshape_cv(x), t_matrix)


def transform_cv2np(x, t_matrix):
    """Apply a transform on a numpy indexed array and return a numpy indexed array."""
    return reshape_cv(cv2.transform(np.array([x]).astype(np.float32), t_matrix)[0])


def rect_scale_pad(rect, scale=1., pad=40.):
    """Scale and/or pad a rectangle."""
    return (rect[0],
            tuple((np.array(rect[1]) + pad) * scale),
            rect[2])


def rect_rotate(rect, angle=None):
    """Rotate a rectangle by an angle in respect to it's center.
    The rect must be of the from (tuple(center_xy), tuple(width_xy), angle).
    The angle is in degrees.
    """
    if angle is None:
        angle = rect[2]
    rad = np.deg2rad(np.abs(angle))
    rot_matrix_2d = np.array([[np.cos(rad), np.sin(rad)],
                              [np.sin(rad), np.cos(rad)]])

    # cal. center of rectangle
    center = np.sum(np.array(rect[1]).reshape(1, -1) * rot_matrix_2d, axis=-1) * .5
    center = np.abs(center)

    return tuple(center), rect[1], angle

例子：

# Generate Image
img = np.zeros((1200, 660), dtype=np.uint8)

# Draw some lines and gen. points
x_0 = np.array([150,600])
x_1 = np.int0(x_0 + np.array((100, 100)))
x_2 = np.int0(x_0 + np.array((100, -100))*2.5)
img = cv2.line(img,tuple(x_0),tuple(x_1),1,120)
img = cv2.line(img,tuple(x_0),tuple(x_2),1,120)
points = np.array([x_0, x_1, x_2])

# Get Box
rect = cv2.minAreaRect(np.argwhere(img))

# Apply transformation
rect_scale = rect_scale_pad(rect, scale = 1., pad = 40.)
img_return, rect_target, t_matrix = img_rectangle_cut(
    img, 
    rect_scale, 
    angle=0,
    angle_normalize=True  # True <-> angel=0 vertical; angel=90 horizontal
   )

# PLOT
fig, ax = plt.subplots(ncols=2, figsize=(10,5))
ax = ax.flatten()
ax[0].imshow(img)

box_i = reshape_cv(cv2.boxPoints(rect))
ax[0].plot(*connect(box_i).T, 'o-', color='gray', alpha=.75, label='Original Box')
box_i = reshape_cv(cv2.boxPoints(rect_scale))
ax[0].plot(*connect(box_i).T, 'o-', color='green', alpha=.75, label='Scaled Box')
ax[0].plot(*points.T, 'o', label='Points')


ax[1].imshow(img_return)
box_i = transform_cv2np(cv2.boxPoints(rect), t_matrix)
ax[1].plot(*connect(box_i).T, 'o-', color='gray', alpha=.75, label='Original Box')

point_t = transform_np(points, t_matrix)
ax[1].plot(*point_t.T, 'o', label='Points')

ax[0].set_title('Original')
ax[1].set_title('Translated')

for axi in ax:
    axi.legend(loc=1)
    
plt.tight_layout()