对齐和裁剪相同的场景图像
Aligning and cropping same scene images
您好，我使用包围曝光拍摄了不同的图像（同一场景不同曝光），我需要对齐图像并裁剪每一张图像，以便它们完全匹配。 （由于拍摄这些照片时有相机抖动）
我不想合并它们，我只想剪切、旋转或缩放 .. 等等，以便它们完全对齐，然后保存它们。如果我知道如何做到这一点，我会添加一个代码示例。但我不知道。我是 opencv 的新手。
举个例子：
这是一个真实的样本示例：（这个样本有很大的错位，大多数样本只需要很小的调整，因为不像这个样本那样摇晃）
我需要裁剪每一张图片，使它们完全相同（只保留共享区域）
谢谢！
关于这个答案，我描述了一种实现 图像对齐 的方法，它包括使用 欧几里得模型 来转换 图像 1（左侧）和 图像 3（右侧）根据 图像 2，中心图像。
但是，我想快速指出，共享的图像非常具有挑战性：不仅对比度差异很大，而且翻译也有显着差异， 缩放、旋转，甚至可能还有一点剪切。事实上，它们以非常低的分辨率发布也无济于事。
无论如何，我指出了它们在 2D 变换 方面的差异，因为在选择合适的 模型[=65] 时，您始终需要牢记这一点=] 执行图像对齐。 This nice picture is from a tutorial that describes the models in greater detail:
该方法包含以下步骤：
- 使用增强相关系数 (ECC) 算法在图像 2 和图像 1 之间使用欧氏模型执行图像对齐；
- 然后利用返回的变换矩阵，借助
cv2.warpAffine()对图像1进行变换，并计算图像1中变换的近似矩形面积；
- 重复相同的步骤对图像 3 进行变换：使用增强相关系数 (ECC) 算法在图像 2 和图像 3 之间使用欧氏模型执行图像对齐；
- 然后使用返回的变换矩阵借助
cv2.warpAffine()对图像3进行变换，并计算出变换的近似矩形面积。
- 这些操作的结果是对齐的图像，可以在下图中看到。绿色矩形显示转换区域：
中间图像中的红色矩形，用于创建转换模型的参考图像，是图像 1 和 3 区域之间的交集，可以看作是公共区域在所有 3 张图像之间。
然后可以使用红色矩形裁剪图像 1、2 和 3，并使图像对齐看起来很漂亮。请注意所有这些图像上的地形和天空如何完美地相互对齐：
注意到这种方法的成本很有趣：因为图像 1 没有捕捉到图像 2 和 3 上可以轻松看到的所有地形特征，最终结果是图像 2 和 3 最终失去了那些地形的一部分。因此，所有图像都显示了照片的完全相同区域。
Python源代码:
###
# reference:
# https://www.learnopencv.com/image-alignment-ecc-in-opencv-c-python/
###
import numpy as np
import cv2
# internalRect: returns the intersection between two rectangles
#
# p1 ---------------- p2
# | |
# | |
# | |
# p4 ---------------- p3
def internalRect(r1, r2):
x = 0
y = 1
w = 2
h = 3
rect1_pt1 = [ r1[x], r1[y] ]
rect1_pt2 = [ r1[x]+r1[w], r1[y] ]
rect1_pt3 = [ r1[x]+r1[w], r1[y]+r1[h] ]
rect1_pt4 = [ r1[x], r1[y]+r1[h] ]
rect2_pt1 = [ r2[x], r2[y] ]
rect2_pt2 = [ r2[x]+r2[w], r2[y] ]
rect2_pt3 = [ r2[x]+r2[w], r2[y]+r2[h] ]
rect2_pt4 = [ r2[x], r2[y]+r2[h] ]
int_pt1 = [ max(rect1_pt1[x], rect2_pt1[x]), max(rect1_pt1[y], rect2_pt1[y]) ]
int_pt2 = [ min(rect1_pt2[x], rect2_pt2[x]), max(rect1_pt2[y], rect2_pt2[y]) ]
int_pt3 = [ min(rect1_pt3[x], rect2_pt3[x]), min(rect1_pt3[y], rect2_pt3[y]) ]
int_pt4 = [ max(rect1_pt4[x], rect2_pt4[x]), min(rect1_pt4[y], rect2_pt4[y]) ]
rect = [ int_pt1[x], int_pt1[y], int_pt2[x]-int_pt1[x], int_pt4[y]-int_pt1[y] ]
return rect
# align_image: use src1 as the reference image to transform src2
def align_image(src1, src2, warp_mode=cv2.MOTION_TRANSLATION):
# convert images to grayscale
img1_gray = cv2.cvtColor(src1, cv2.COLOR_BGR2GRAY)
img2_gray = cv2.cvtColor(src2, cv2.COLOR_BGR2GRAY)
# define 2x3 or 3x3 matrices and initialize it to a identity matrix
if warp_mode == cv2.MOTION_HOMOGRAPHY:
warp_matrix = np.eye(3, 3, dtype=np.float32)
else:
warp_matrix = np.eye(2, 3, dtype=np.float32)
# number of iterations:
num_iters = 1000
# specify the threshold of the increment in the correlation coefficient between two iterations
termination_eps = 1e-8
# Define termination criteria
criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, num_iters, termination_eps)
print('findTransformECC() may take a while...')
# perform ECC: use the selected model to calculate the transformation required to align src2 with src1. The resulting transformation matrix is stored in warp_matrix:
(cc, warp_matrix) = cv2.findTransformECC(img1_gray, img2_gray, warp_matrix, warp_mode, criteria, inputMask=None, gaussFiltSize=1)
if (warp_mode == cv2.MOTION_HOMOGRAPHY):
img2_aligned = cv2.warpPerspective(src2, warp_matrix, (src1.shape[1], src1.shape[0]), flags=cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP)
else :
# use warpAffine() for: translation, euclidean and affine models
img2_aligned = cv2.warpAffine(src2, warp_matrix, (src1.shape[1], src1.shape[0]), flags=cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP, borderMode=cv2.BORDER_CONSTANT, borderValue=0)
#print('warp_matrix shape', warp_matrix.shape, 'data=\n', warp_matrix)
#print(warp_matrix, warp_matrix)
# compute the cropping area to remove the black bars from the transformed image
x = 0
y = 0
w = src1.shape[1]
h = src1.shape[0]
if (warp_matrix[0][2] < 0):
x = warp_matrix[0][2] * -1
w -= x
if (warp_matrix[1][2] < 0):
y = warp_matrix[1][2] * -1
h -= y
if (warp_matrix[1][2] > 0):
h -= warp_matrix[1][2]
matchArea = [ int(x), int(y), int(w), int(h) ]
#print('src1 w=', src1.shape[1], 'h=', src1.shape[0])
#print('matchedRect=', matchArea[0], ',', matchArea[1], '@', matchArea[2], 'x', matchArea[3], '\n')
return img2_aligned, matchArea
##########################################################################################
img1 = cv2.imread("img1.png")
img2 = cv2.imread("img2.png")
img3 = cv2.imread("img3.png")
# TODO: adjust contrast on all input images
###
# resize images to be the same size as the smallest image for debug purposes
###
max_h = img1.shape[0]
max_h = max(max_h, img2.shape[0])
max_h = max(max_h, img3.shape[0])
max_w = img1.shape[1]
max_w = max(max_w, img2.shape[1])
max_w = max(max_w, img3.shape[1])
img1_padded = cv2.resize(img1, (max_w, max_h), interpolation=cv2.INTER_AREA)
img2_padded = cv2.resize(img2, (max_w, max_h), interpolation=cv2.INTER_AREA)
img3_padded = cv2.resize(img3, (max_w, max_h), interpolation=cv2.INTER_AREA)
# stack them horizontally for display
hStack = np.hstack((img1_padded, img2_padded)) # stack images side-by-side
input_stacked = np.hstack((hStack, img3_padded)) # stack images side-by-side
cv2.imwrite("input_stacked.jpg", input_stacked)
cv2.imshow("input_stacked", input_stacked)
cv2.waitKey(0)
###
# perform image alignment
###
# specify the motion model
warp_mode = cv2.MOTION_EUCLIDEAN # cv2.MOTION_TRANSLATION, cv2.MOTION_EUCLIDEAN, cv2.MOTION_AFFINE, cv2.MOTION_HOMOGRAPHY
# for testing purposes: img2 will be the reference image
img1_aligned, matchArea1 = align_image(img2, img1, warp_mode)
img1_aligned_cpy = img1_aligned.copy()
cv2.rectangle(img1_aligned_cpy, (matchArea1[0], matchArea1[1]), (matchArea1[0]+matchArea1[2], matchArea1[1]+matchArea1[3]), (0, 255, 0), 2)
cv2.imwrite("img1_aligned.jpg", img1_aligned_cpy)
print('\n###############################################\n')
# for testing purposes: img2 will be the reference image again
img3_aligned, matchArea3 = align_image(img2, img3, warp_mode)
img3_aligned_cpy = img3_aligned.copy()
cv2.rectangle(img3_aligned_cpy, (matchArea3[0], matchArea3[1]), (matchArea3[0]+matchArea3[2], matchArea3[1]+matchArea3[3]), (0, 255, 0), 2)
cv2.imwrite("img3_aligned.jpg", img3_aligned_cpy)
# compute the crop area in the reference image and draw a red rectangle
cropRect = internalRect(matchArea1, matchArea3)
print('cropRect=', cropRect[0], ',', cropRect[1], '@', cropRect[2], 'x', cropRect[3], '\n')
img2_eq_cpy = img2.copy()
cv2.rectangle(img2_eq_cpy, (cropRect[0], cropRect[1]), (cropRect[0]+cropRect[2], cropRect[1]+cropRect[3]), (0, 0, 255), 2)
cv2.imwrite("img2_eq.jpg", img2_eq_cpy)
# stack results horizontally for display
res_hStack = np.hstack((img1_aligned_cpy, img2_eq_cpy)) # stack images side-by-side
aligned_stacked = np.hstack((res_hStack, img3_aligned_cpy)) # stack images side-by-side
cv2.imwrite("aligned_stacked.jpg", aligned_stacked)
cv2.imshow("aligned_stacked", aligned_stacked)
cv2.waitKey(0)
print('\n###############################################\n')
# crop images to the smallest internal area between them
img1_aligned_cropped = img1_aligned[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
img3_aligned_cropped = img3_aligned[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
img2_eq_cropped = img2[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
cropped_hStack = np.hstack((img1_aligned_cropped, img2_eq_cropped)) # stack images side-by-side
cropped_stacked = np.hstack((cropped_hStack, img3_aligned_cropped)) # stack images side-by-side
cv2.imwrite("cropped_stacked.jpg", cropped_stacked)
cv2.imshow("cropped_stacked", cropped_stacked)
cv2.waitKey(0)
您好，我使用包围曝光拍摄了不同的图像（同一场景不同曝光），我需要对齐图像并裁剪每一张图像，以便它们完全匹配。 （由于拍摄这些照片时有相机抖动）
我不想合并它们，我只想剪切、旋转或缩放 .. 等等，以便它们完全对齐，然后保存它们。如果我知道如何做到这一点，我会添加一个代码示例。但我不知道。我是 opencv 的新手。
举个例子：
这是一个真实的样本示例：（这个样本有很大的错位，大多数样本只需要很小的调整，因为不像这个样本那样摇晃）
我需要裁剪每一张图片，使它们完全相同（只保留共享区域）
谢谢！
关于这个答案，我描述了一种实现 图像对齐 的方法，它包括使用 欧几里得模型 来转换 图像 1（左侧）和 图像 3（右侧）根据 图像 2，中心图像。
但是，我想快速指出，共享的图像非常具有挑战性：不仅对比度差异很大，而且翻译也有显着差异， 缩放、旋转，甚至可能还有一点剪切。事实上，它们以非常低的分辨率发布也无济于事。
无论如何，我指出了它们在 2D 变换 方面的差异，因为在选择合适的 模型[=65] 时，您始终需要牢记这一点=] 执行图像对齐。 This nice picture is from a tutorial that describes the models in greater detail:
该方法包含以下步骤：
- 使用增强相关系数 (ECC) 算法在图像 2 和图像 1 之间使用欧氏模型执行图像对齐；
- 然后利用返回的变换矩阵，借助
cv2.warpAffine()对图像1进行变换，并计算图像1中变换的近似矩形面积； - 重复相同的步骤对图像 3 进行变换：使用增强相关系数 (ECC) 算法在图像 2 和图像 3 之间使用欧氏模型执行图像对齐；
- 然后使用返回的变换矩阵借助
cv2.warpAffine()对图像3进行变换，并计算出变换的近似矩形面积。 - 这些操作的结果是对齐的图像，可以在下图中看到。绿色矩形显示转换区域：
中间图像中的红色矩形，用于创建转换模型的参考图像，是图像 1 和 3 区域之间的交集，可以看作是公共区域在所有 3 张图像之间。
然后可以使用红色矩形裁剪图像 1、2 和 3，并使图像对齐看起来很漂亮。请注意所有这些图像上的地形和天空如何完美地相互对齐：
注意到这种方法的成本很有趣：因为图像 1 没有捕捉到图像 2 和 3 上可以轻松看到的所有地形特征，最终结果是图像 2 和 3 最终失去了那些地形的一部分。因此，所有图像都显示了照片的完全相同区域。
Python源代码:
###
# reference:
# https://www.learnopencv.com/image-alignment-ecc-in-opencv-c-python/
###
import numpy as np
import cv2
# internalRect: returns the intersection between two rectangles
#
# p1 ---------------- p2
# | |
# | |
# | |
# p4 ---------------- p3
def internalRect(r1, r2):
x = 0
y = 1
w = 2
h = 3
rect1_pt1 = [ r1[x], r1[y] ]
rect1_pt2 = [ r1[x]+r1[w], r1[y] ]
rect1_pt3 = [ r1[x]+r1[w], r1[y]+r1[h] ]
rect1_pt4 = [ r1[x], r1[y]+r1[h] ]
rect2_pt1 = [ r2[x], r2[y] ]
rect2_pt2 = [ r2[x]+r2[w], r2[y] ]
rect2_pt3 = [ r2[x]+r2[w], r2[y]+r2[h] ]
rect2_pt4 = [ r2[x], r2[y]+r2[h] ]
int_pt1 = [ max(rect1_pt1[x], rect2_pt1[x]), max(rect1_pt1[y], rect2_pt1[y]) ]
int_pt2 = [ min(rect1_pt2[x], rect2_pt2[x]), max(rect1_pt2[y], rect2_pt2[y]) ]
int_pt3 = [ min(rect1_pt3[x], rect2_pt3[x]), min(rect1_pt3[y], rect2_pt3[y]) ]
int_pt4 = [ max(rect1_pt4[x], rect2_pt4[x]), min(rect1_pt4[y], rect2_pt4[y]) ]
rect = [ int_pt1[x], int_pt1[y], int_pt2[x]-int_pt1[x], int_pt4[y]-int_pt1[y] ]
return rect
# align_image: use src1 as the reference image to transform src2
def align_image(src1, src2, warp_mode=cv2.MOTION_TRANSLATION):
# convert images to grayscale
img1_gray = cv2.cvtColor(src1, cv2.COLOR_BGR2GRAY)
img2_gray = cv2.cvtColor(src2, cv2.COLOR_BGR2GRAY)
# define 2x3 or 3x3 matrices and initialize it to a identity matrix
if warp_mode == cv2.MOTION_HOMOGRAPHY:
warp_matrix = np.eye(3, 3, dtype=np.float32)
else:
warp_matrix = np.eye(2, 3, dtype=np.float32)
# number of iterations:
num_iters = 1000
# specify the threshold of the increment in the correlation coefficient between two iterations
termination_eps = 1e-8
# Define termination criteria
criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, num_iters, termination_eps)
print('findTransformECC() may take a while...')
# perform ECC: use the selected model to calculate the transformation required to align src2 with src1. The resulting transformation matrix is stored in warp_matrix:
(cc, warp_matrix) = cv2.findTransformECC(img1_gray, img2_gray, warp_matrix, warp_mode, criteria, inputMask=None, gaussFiltSize=1)
if (warp_mode == cv2.MOTION_HOMOGRAPHY):
img2_aligned = cv2.warpPerspective(src2, warp_matrix, (src1.shape[1], src1.shape[0]), flags=cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP)
else :
# use warpAffine() for: translation, euclidean and affine models
img2_aligned = cv2.warpAffine(src2, warp_matrix, (src1.shape[1], src1.shape[0]), flags=cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP, borderMode=cv2.BORDER_CONSTANT, borderValue=0)
#print('warp_matrix shape', warp_matrix.shape, 'data=\n', warp_matrix)
#print(warp_matrix, warp_matrix)
# compute the cropping area to remove the black bars from the transformed image
x = 0
y = 0
w = src1.shape[1]
h = src1.shape[0]
if (warp_matrix[0][2] < 0):
x = warp_matrix[0][2] * -1
w -= x
if (warp_matrix[1][2] < 0):
y = warp_matrix[1][2] * -1
h -= y
if (warp_matrix[1][2] > 0):
h -= warp_matrix[1][2]
matchArea = [ int(x), int(y), int(w), int(h) ]
#print('src1 w=', src1.shape[1], 'h=', src1.shape[0])
#print('matchedRect=', matchArea[0], ',', matchArea[1], '@', matchArea[2], 'x', matchArea[3], '\n')
return img2_aligned, matchArea
##########################################################################################
img1 = cv2.imread("img1.png")
img2 = cv2.imread("img2.png")
img3 = cv2.imread("img3.png")
# TODO: adjust contrast on all input images
###
# resize images to be the same size as the smallest image for debug purposes
###
max_h = img1.shape[0]
max_h = max(max_h, img2.shape[0])
max_h = max(max_h, img3.shape[0])
max_w = img1.shape[1]
max_w = max(max_w, img2.shape[1])
max_w = max(max_w, img3.shape[1])
img1_padded = cv2.resize(img1, (max_w, max_h), interpolation=cv2.INTER_AREA)
img2_padded = cv2.resize(img2, (max_w, max_h), interpolation=cv2.INTER_AREA)
img3_padded = cv2.resize(img3, (max_w, max_h), interpolation=cv2.INTER_AREA)
# stack them horizontally for display
hStack = np.hstack((img1_padded, img2_padded)) # stack images side-by-side
input_stacked = np.hstack((hStack, img3_padded)) # stack images side-by-side
cv2.imwrite("input_stacked.jpg", input_stacked)
cv2.imshow("input_stacked", input_stacked)
cv2.waitKey(0)
###
# perform image alignment
###
# specify the motion model
warp_mode = cv2.MOTION_EUCLIDEAN # cv2.MOTION_TRANSLATION, cv2.MOTION_EUCLIDEAN, cv2.MOTION_AFFINE, cv2.MOTION_HOMOGRAPHY
# for testing purposes: img2 will be the reference image
img1_aligned, matchArea1 = align_image(img2, img1, warp_mode)
img1_aligned_cpy = img1_aligned.copy()
cv2.rectangle(img1_aligned_cpy, (matchArea1[0], matchArea1[1]), (matchArea1[0]+matchArea1[2], matchArea1[1]+matchArea1[3]), (0, 255, 0), 2)
cv2.imwrite("img1_aligned.jpg", img1_aligned_cpy)
print('\n###############################################\n')
# for testing purposes: img2 will be the reference image again
img3_aligned, matchArea3 = align_image(img2, img3, warp_mode)
img3_aligned_cpy = img3_aligned.copy()
cv2.rectangle(img3_aligned_cpy, (matchArea3[0], matchArea3[1]), (matchArea3[0]+matchArea3[2], matchArea3[1]+matchArea3[3]), (0, 255, 0), 2)
cv2.imwrite("img3_aligned.jpg", img3_aligned_cpy)
# compute the crop area in the reference image and draw a red rectangle
cropRect = internalRect(matchArea1, matchArea3)
print('cropRect=', cropRect[0], ',', cropRect[1], '@', cropRect[2], 'x', cropRect[3], '\n')
img2_eq_cpy = img2.copy()
cv2.rectangle(img2_eq_cpy, (cropRect[0], cropRect[1]), (cropRect[0]+cropRect[2], cropRect[1]+cropRect[3]), (0, 0, 255), 2)
cv2.imwrite("img2_eq.jpg", img2_eq_cpy)
# stack results horizontally for display
res_hStack = np.hstack((img1_aligned_cpy, img2_eq_cpy)) # stack images side-by-side
aligned_stacked = np.hstack((res_hStack, img3_aligned_cpy)) # stack images side-by-side
cv2.imwrite("aligned_stacked.jpg", aligned_stacked)
cv2.imshow("aligned_stacked", aligned_stacked)
cv2.waitKey(0)
print('\n###############################################\n')
# crop images to the smallest internal area between them
img1_aligned_cropped = img1_aligned[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
img3_aligned_cropped = img3_aligned[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
img2_eq_cropped = img2[cropRect[1] : cropRect[1]+cropRect[3], cropRect[0] : cropRect[0]+cropRect[2]]
cropped_hStack = np.hstack((img1_aligned_cropped, img2_eq_cropped)) # stack images side-by-side
cropped_stacked = np.hstack((cropped_hStack, img3_aligned_cropped)) # stack images side-by-side
cv2.imwrite("cropped_stacked.jpg", cropped_stacked)
cv2.imshow("cropped_stacked", cropped_stacked)
cv2.waitKey(0)