24. image mosaic operation

1. Basic knowledge of image mosaic

  Prepare two pieces. Related Pictures of relationships

Homework after class ： Remove the edges where two pictures meet , Crop the image to remove the black edge

（1） Steps of image merging

1. Read the file and reset the size
2. Calculate descriptors based on feature points and , The homography matrix is obtained
3. Image transformation
4. Image mosaic and output image

Find the center point first

#  Image mosaic operation 
import cv2
import numpy as np

#  First step , Read the file , Set the picture to the same size ,640*480（ The size closest to the actual image ）
img1 = cv2.imread('E:\\opencv_photo\\map1.png')
img2 = cv2.imread('E:\\opencv_photo\\map2.png')
img1 = cv2.resize(img1,(610,480))
img2 = cv2.resize(img2,(610,480))
#  Arrange the two figures together 
inputs = np.hstack((img1, img2))
#  The second step , Finding feature points , Computational descriptors , Calculate the homography matrix 
#  The third step , The image is transformed according to the homography matrix , And then translation 
#  Step four , Splice and output the final result 

cv2.imshow('input img',inputs)
cv2.waitKey()

#  Image mosaic operation 
import cv2
import numpy as np

def stitch_image(img1, img2, H) :
    # 1. Get... For each picture 4 Corner points 
    # 2. Transform the picture , The image is rotated by homography matrix , Then pan 
    # 3. Create a large picture , Join the two figures together 
    # 4. Output the result 

    h1, w1 = img1.shape[:2]     #  Get the height and width of the original graph 
    h2, w2 = img2.shape[:2]
    img1_dims = np.float32([[0,0],[0,h1],[w1,h1],[w1,0]]).reshape(-1,1,2) #  The four corners are sorted counterclockwise 
    img2_dims = np.float32([[0,0],[0,h2],[w2,h2],[w2,0]]).reshape(-1,1,2)
    #  Image transformation 
    img1_transform = cv2.perspectiveTransform(img1_dims,H)
    # print(img1_dims)
    # print(img2_dims)
    # print(img1_transform)
    result_dims = np.concatenate((img2_dims, img1_transform), axis=0)   #  Horizontal splicing , It is mainly to find the maximum and minimum value of the image 
    # print(result_dims)    #  The result of mixing 

    [x_min, y_min] = np.int32(result_dims.min(axis=0).ravel()-0.5)  # axis=0 Press x Axis get value ,ravel() Change two dimensions into one dimension , According to the principle of rounding , minimum value -0.5, Maximum +0.5
    [x_max, y_max ] = np.int32(result_dims.max(axis=0).ravel()+0.5)
    #  The distance of translation 
    transform_dist = [-x_min, -y_min]  #  The smallest x Distance and y

    #[1, 0, dx]
    #[0, 1, dy]         
    #[0, 0, 1 ]
    #  The method of matrix translation , Multiply by a homogeneous matrix 
    transform_array = np.array([[1, 0, transform_dist[0]],
                                [0, 1, transform_dist[1]],
                                [0, 0, 1]])
    #  Projection transformation 
    result_img = cv2.warpPerspective(img1, transform_array.dot(H), (x_max-x_min, y_max-y_min))

    result_img[transform_dist[1]:transform_dist[1]+h2, 
                transform_dist[0]:transform_dist[0]+w2] = img2

    return result_img

#  The method of calculating homography matrix 
def get_homo(img1, img2):
    # 1. Create feature transformation objects 
    sift = cv2.xfeatures2d.SIFT_create()
    # 2. The feature points and descriptors are obtained through the feature transformation object 
    k1,d1 = sift.detectAndCompute(img1,None)    # k1 For feature points ,d1 Is a descriptor 
    k2,d2 = sift.detectAndCompute(img2,None)
    # 3. Create feature matching 
    # 4. Feature matching 
    # 5. Filtering characteristics , Find out the effective feature matching points 
    bf = cv2.BFMatcher()  #  Use violence feature matching 
    matches = bf.knnMatch(d1,d2,k=2)    #  Matching feature points 
    
    verify_ratio = 0.8  #  Filter 
    verify_matches = []
    for m1, m2 in matches:  # m1,m2 The smaller the distance, the better , Less than 0.8 Consider effective , Greater than 0.8 Invalid 
        if m1.distance < 0.8*m2.distance:
            verify_matches.append(m1)

    #  Minimum number of matches 
    min_matches = 8
    if len(verify_matches) > min_matches:
        img1_pts = []
        img2_pts = []
        for m in verify_matches:
            img1_pts.append(k1[m.queryIdx].pt)  #  Images 1 Coordinate feature points of 
            img2_pts.append(k2[m.trainIdx].pt)  #  Images 2 Coordinate feature points of 
        #[(x1, y1), (x2, y2), ...]
        #[[x1, y1], [x2, y2], ...]

        img1_pts = np.float32(img1_pts).reshape(-1,1,2)    #  To adapt to findHomography The format of 
        img2_pts = np.float32(img2_pts).reshape(-1,1,2) 
        H, mask =  cv2.findHomography(img1_pts, img2_pts, cv2.RANSAC,5.0)  #  Get homography matrix 
        return H
    else:
        print('error: Not enough matches!')
        exit()
    
#  First step , Read the file , Set the picture to the same size ,640*480（ The size closest to the actual image ）
img1 = cv2.imread('E:\\opencv_photo\\map1.png')
img2 = cv2.imread('E:\\opencv_photo\\map2.png')
img1 = cv2.resize(img1,(640,480))
img2 = cv2.resize(img2,(640,480))
#  Arrange the two figures together 
inputs = np.hstack((img1, img2))

#  The second step , Finding feature points , Computational descriptors , Calculate the homography matrix 
H = get_homo(img1, img2)    #  Obtain the homography matrix 

#  The third step , The image is transformed according to the homography matrix , And then translation 
result_image = stitch_image(img1, img2, H)      #  Image mosaic 
#  Step four , Splice and output the final result 

cv2.imshow('input img',result_image)
cv2.waitKey()