python中的圖像註冊

Question

我正在研究一個項目，該項目需要查找新圖像移動並旋轉w.r.t舊圖像的程度。我正在嘗試使用fft來實現它。但是，它適用於某些情況，而其他情況不適用。我遵循的步驟如下：

1. 採取兩個圖像，實現Canny邊緣檢測
2. 使用FFT找出移動和刪除移位
3. 圖像變換極性域名，並找到移動和轉換弧度回答

在某些情況下，我可以得到正確的答案，但在其他情況下，我的換檔角度爲（0,0），旋轉角度爲0弧度。請提出可能發生這種情況的原因。

下面是代碼：

class Register: 
    ''' 
     Class for registering images based on FFT. The usage is as follows: 
     >>> im0 = imread('image0.jpg', flatten = True) 
     >>> im1 = imread('image1.jpg', flatten = True) 
     >>> reg = Register(im0, im1) 
     >>> shift = reg.shift 
     >>> rotation = reg.theta 

     Note: 
      1. This image registration technique is not very reliable and 
       is valid only for small rotation 
      2. The class is very slow, since it depends on canny edge detection 
       module for finding edges. 
    ''' 
    def __init__(self,imin0, imin1, PROCESSED = False): 
     ''' 
      This method is used to execute all the routines required to get the 
      shift and the rotation 
     ''' 
     # find edges to remove low frequency signals and suppress information 
     if PROCESSED: 
      im0 = imin0 
      im1 = imin1 
     else: 
      im0 = Canny(imin0, 0.85, 5).grad 
      im1 = Canny(imin1, 0.85, 5).grad 

     # A major drawback of this method is that it can operate only on square 
     # images. Hence we will make square image of any input image 

     im0 = self.createsquareim(self.clearBorder(im0)) 
     im1 = self.createsquareim(self.clearBorder(im1)) 

     self.shift = self.findShift(im0,im1) 
     imtrans = shift(im1, self.shift) 
     # Remove the shift in the image. This is mandatory before we find theta 
     impolar0 = self.makePolar(im0) 
     impolar1 = self.makePolar(imtrans) 
     self.index = self.findShift(impolar0, impolar1)[1] 
     self.theta = ((self.index*90.0)/impolar1.shape[0]) 

    def clearBorder(self,im,width = 50, color = 255): 
     ''' 
      A little house keeping to clear any border noise 
     ''' 
     im[:,:width] = color 
     im[:,-width:] = color 
     im[:width,:] = color 
     im[-width:,:] = color 

     return im 

    def createsquareim(self, im): 
     """ 
     function createsquareim 
     input:numpy ndarray 
     output:numpy ndarray 

     The function takes in an image array and converts it into square 
     image by creating empty columns and rows. 
     """ 
     lenmax = max(im.shape[0],im.shape[1]) 
     imout = zeros((lenmax,lenmax)) 
     imout[:,:] = 255 
     imout[:im.shape[0],:im.shape[1]] = im 
     return imout 

    def findShift(self, im0, im1): 
     ''' 
      This method is based on fft method of registering images. 
     ''' 
     IM0 = fft2(im0) 
     IM1 = fft2(im1) 

     numer = IM0*conj(IM1) 
     denom = abs(IM0*IM1) 

     pulse_im = ifft2(numer/denom) 
     mag = abs(pulse_im) 
     x, y = where(mag == mag.max()) 

     x = array(x.tolist()) # Issues with read only arrays 
     y = array(y.tolist()) 

     X, Y = im0.shape 

     if x > X/2: 
      x -= X 
     if y > Y/2: 
      y -= Y 

     return [x[0], y[0]] 

    def makePolar(self, im): 
     ''' 
      This method will convert the cartesian coordinates image 
      to polar coordinates image. The relation between the two 
      domains is 
       F(r,theta) = f(r*cos(theta),r*sin(theta)) 
      To make the process fast, we are using map_coordinates function 
     ''' 
     m, n = im.shape 
     r_max = hypot(m, n) 

     r_mat = zeros_like(im) 
     t_mat = zeros_like(im) 

     r_mat.T[:] = linspace(0, r_max, m) 
     t_mat[:] = linspace(0, pi/2, n) 

     x = r_mat*cos(t_mat) 
     y = r_mat*sin(t_mat) 

     imout = zeros_like(im) 
     imout = map_coordinates(im, [x, y], cval = 255) 

     return imout 

Answer 1

我已經寫了使用SIFT和RANSAC算法圖像配準Python腳本，你可以在http://github.com/vishwa91/pyimreg找到的代碼。另外，同樣的小介紹可以在http://cyroforge.wordpress.com/2012/07/05/image-registration-using-python/