一、前言

图像配准是一种图像处理技术，用于将多个场景对齐到单个集成图像中。在这篇文章中，我将讨论如何在可见光及其相应的热图像上应用图像配准。在继续该过程之前，让我们看看什么是热图像及其属性。

二、热红外数据介绍

热图像本质上通常是灰度图像：黑色物体是冷的，白色物体是热的，灰色的深度表示两者之间的差异。 然而，一些热像仪会为图像添加颜色，以帮助用户识别不同温度下的物体。

图1 左图为可见光；有图为热红外图像

上面两个图像是可见的，它是对应的热图像，你可以看到热图像有点被裁剪掉了。 这是因为在热图像中并没有捕获整个场景，而是将额外的细节作为元数据存储在热图像中。

因此，为了执行配准，我们要做的是找出可见图像的哪一部分出现在热图像中，然后对图像的该部分应用配准。

图2 .与热图像匹配后裁剪的可见图像

为了执行上述操作，基本上包含两张图像，一张参考图像和另一张要匹配的图像。 因此，下面的算法会找出参考图像的哪一部分出现在第二张图像中，并为您提供匹配图像部分的位置。

现在我们知道热图像中存在可见图像的哪一部分，我们可以裁剪可见图像，然后对生成的图像进行配准。

三、配准过程

为了执行配准，我们要做的是找出将像素从可见图像映射到热图像的特征点，这在本文中进行了解释，一旦我们获得了一定数量的像素，我们就会停止并开始映射这些像素，从而完成配准过程完成了。

图3 热成像到可见光图像配准

一旦我们执行了配准，如果匹配正确，我们将获得具有配准图像的输出，如下图所示。

图4 最终输出结果

我对 400 张图像的数据集执行了此操作，获得的结果非常好。 错误数量很少，请参考下面的代码，看看一切是如何完成的。

from __future__ import print_functionimport numpy as npimport argparseimport imutilsimport globimport cv2import osMAX_FEATURES = 500GOOD_MATCH_PERCENT = 0.15#function to align the thermal and visible image, it returns the homography matrix def alignImages(im1, im2,filename): # Convert images to grayscaleim1Gray = cv2.cvtColor(im1, cv2.COLOR_BGR2GRAY)im2Gray = cv2.cvtColor(im2, cv2.COLOR_BGR2GRAY) # Detect ORB features and compute descriptors.orb = cv2.ORB_create(MAX_FEATURES)keypoints1, descriptors1 = orb.detectAndCompute(im1Gray, None)keypoints2, descriptors2 = orb.detectAndCompute(im2Gray, None) # Match features.matcher = cv2.DescriptorMatcher_create(cv2.DESCRIPTOR_MATCHER_BRUTEFORCE_HAMMING)matches = matcher.match(descriptors1, descriptors2, None) # Sort matches by scorematches.sort(key=lambda x: x.distance, reverse=False) # Remove not so good matchesnumGoodMatches = int(len(matches) * GOOD_MATCH_PERCENT)matches = matches[:numGoodMatches] # Draw top matchesimMatches = cv2.drawMatches(im1, keypoints1, im2, keypoints2, matches, None)cv2.imwrite(os.path.join('./registration/',filename), imMatches) # Extract location of good matchespoints1 = np.zeros((len(matches), 2), dtype=np.float32)points2 = np.zeros((len(matches), 2), dtype=np.float32) for i, match in enumerate(matches):points1[i, :] = keypoints1[match.queryIdx].ptpoints2[i, :] = keypoints2[match.trainIdx].pt # Find homographyh, mask = cv2.findHomography(points1, points2, cv2.RANSAC) # Use homographyheight, width, channels = im2.shapeim1Reg = cv2.warpPerspective(im1, h, (width, height)) return im1Reg, h# construct the argument parser and parse the arguments# run the file with python registration.py --image filenameap = argparse.ArgumentParser()# ap.add_argument("-t", "--template", required=True, help="Path to template image")ap.add_argument("-i", "--image", required=True,help="Path to images where template will be matched")ap.add_argument("-v", "--visualize",help="Flag indicating whether or not to visualize each iteration")args = vars(ap.parse_args())# put the thermal image in a folder named thermal and the visible image in a folder named visible with the same name# load the image image, convert it to grayscale, and detect edgestemplate = cv2.imread('thermal/'+args["image"]+'.jpg')template = cv2.cvtColor(template, cv2.COLOR_BGR2GRAY)template = cv2.Canny(template, 50, 200)(tH, tW) = template.shape[:2]cv2.imshow("Template", template)# loop over the images to find the template in# load the image, convert it to grayscale, and initialize the# bookkeeping variable to keep track of the matched regionimage = cv2.imread('visible/'+args["image"]+'.jpg')gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)found = None# loop over the scales of the imagefor scale in np.linspace(0.2, 1.0, 20)[::-1]:# resize the image according to the scale, and keep track# of the ratio of the resizingresized = imutils.resize(gray, width = int(gray.shape[1] * scale))r = gray.shape[1] / float(resized.shape[1])# if the resized image is smaller than the template, then break# from the loopif resized.shape[0] < tH or resized.shape[1]  found[0]:found = (maxVal, maxLoc, r)# unpack the bookkeeping variable and compute the (x, y) coordinates# of the bounding box based on the resized ratio(_, maxLoc, r) = found(startX, startY) = (int(maxLoc[0] * r), int(maxLoc[1] * r))(endX, endY) = (int((maxLoc[0] + tW) * r), int((maxLoc[1] + tH) * r))# draw a bounding box around the detected result and display the imagecv2.rectangle(image, (startX, startY), (endX, endY), (0, 0, 255), 2)crop_img = image[startY:endY, startX:endX]cv2.imshow("Image", image)cv2.imshow("Image", crop_img)name = "thermal/"+args["image"]+'.jpg'thermal_image = cv2.imread(name, cv2.IMREAD_COLOR)#cropping out the matched part of the thermal imagecrop_img = cv2.resize(crop_img, (thermal_image.shape[1], thermal_image.shape[0]))#cropped image will be saved in a folder named outputcv2.imwrite(os.path.join('./output/', args["image"]+'.jpg'),crop_img)#both images are concatenated and saved in a folder named resultsfinal = np.concatenate((crop_img, thermal_image), axis = 1)cv2.imwrite(os.path.join('./results/', args["image"]+'.jpg'),final)cv2.waitKey(0)# Registration# Read reference imagerefFilename = "thermal/"+args["image"]+'.jpg'print("Reading reference image : ", refFilename)imReference = cv2.imread(refFilename, cv2.IMREAD_COLOR)# Read image to be alignedimFilename = "output/"+args["image"]+'.jpg'print("Reading image to align : ", imFilename);im = cv2.imread(imFilename, cv2.IMREAD_COLOR)file_name=args["image"]+'.jpg'imReg, h = alignImages(im,imReference,file_name)print("Estimated homography : \n",h)

我们已经成功地进行了热到可见图像配准。你可以用你的数据集来尝试一下，然后看看结果。