向后扭曲图像调整大小

编程问答 2022-06-03

问题描述

现在，我正在尝试通过对计算出的矩阵M进行反向扭曲来转换原始图像。但是我不知道如何在反翘曲过程中确定 dst图像的高度和宽度 。这是一些细节。

1。使用SIFT算法在图像1和图像2中找到特征和描述符

enter image description here

image1

enter image description here

image2

2。在方程Ax = B中找到A，B和M矩阵

3。使用双线性插值法执行向后弯曲

4。我想使用向后警告获取的图像如下。

enter image description here

接下来是我的代码。 我应该如何确定高度和宽度？ 

import cv2
import numpy as np
import random


def L2_distance(vector1,vector2):
    distance = np.sqrt(np.sum((vector1 - vector2) ** 2))
    return distance

步骤1：使用SIFT查找特征点和描述符

# <main>

img1 = cv2.imread("../resources/image/building.jpg")
img2 = cv2.imread("../resources/image/building_temp.jpg")


threshold = 300
sift = cv2.xfeatures2d.SIFT_create(0)

kp1,des1 = sift.detectAndCompute(img1,None)
kp2,des2 = sift.detectAndCompute(img2,None)

distance = []
for idx_1,des_1 in enumerate(des1):
    dist = []
    for idx_2,des_2 in enumerate(des2):
        dist.append(L2_distance(des_1,des_2))
    distance.append(dist)


distance = np.array(distance)
min_dist_idx = np.argmin(distance,axis=1)
min_dist_value = np.min(distance,axis=1)

points = []

for idx,point in enumerate(kp1):
    if min_dist_value[idx] >=300:
        continue
    x1,y1 = point.pt
    x2,y2 = kp2[min_dist_idx[idx]].pt

    x1 = int(np.round(x1))
    y1 = int(np.round(y1))

    x2 = int(np.round(x2))
    y2 = int(np.round(y2))
    points.append([(x1,y1),(x2,y2)])

步骤2。得到等式Ax = B中的A，B矩阵和变换矩阵M 

A = []
B = []
for idx,point in enumerate(points):
    A.append(list(point[0]) + [1,0])
    A.append([0,0] + list(point[0]) + [1])
    B.append(point[1][0])
    B.append(point[1][1])

A = np.array(A)
B = np.array(B)


inverse = np.linalg.inv(np.dot(A.T,A))
X = np.dot(inverse,A.T)
X = np.dot(X,B)


M = list(X) + [0,1]
M = np.array(M).reshape((3,3))
# inverse of matrix M
M_ = np.linalg.inv(M)

Step3。我的向后翘曲方法

h,w = img1.shape[:2]
size_vec = np.dot(M,np.array([[h,w,1]]).T)
h = abs(int(size_vec[0,0]))
w = abs(int(size_vec[1,0]))
dst = np.zeros((h,3))
h_,w_ = dst.shape[:2]

for row_ in range(h_):
    for col_ in range(w_):
        # bilinear
        vec = np.dot(M_,np.array([[col_,row_,1]]).T)
        c = vec[0,0]
        r = vec[1,0]
        c_left = int(c)  # 버림
        c_right = min(int(c + 1),w - 1)  # 올림
        r_top = int(r)  # 버림
        r_bottom = min(int(r + 1),h - 1)  # 올림

        s = c - c_left
        t = r - r_top
    """
    transformed coordinates as follow
    for example c = x = 1.7,r = y = 2.3

    c_left = x1 = 1,c_ right = x2 = 2
    r_top = y1 = 2,r_bottom = y2 = 3 

                              |  (y1,x1)             (y1,x2)
                              |  
    (y',x') ---------------->   |         
                              |              (y,x)
                              |  (y2,x1)             (y2,x2)

     <transformed img>               <original img>
    """
    intensity = (1 - s) * (1 - t) * img1[r_top,c_left] \
                + s * (1 - t) * img1[r_top,c_right] \
                + (1 - s) * t * img1[r_bottom,c_left] \
                + s * t * img1[r_bottom,c_right]
    dst[row_,col_] = intensity

dst = dst.astype(np.uint8)

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）

bilinear-interpolationsift