Research On Registration Algorithm Of Infrared And Visible Images In Complex Scenes

Infrared imaging sensors have the advantages of long-range detection,smoke penetration and being able to work all day,but the ability of infrared images to reflect target information is weak.Visible images have high resolution and rich information,but they are susceptible to weather conditions such as illumination.The two images have good complementary characteristics,and a comprehensive analysis of the two images can provide more information.Before integrating the two kinds of images,it’s required to realize stable and reliable image registration.However,the existing registration methods have poor performance in complex scenes such as significant resolution differences,blurred images and low quality,and its poor robustness is difficult to meet the wide application requirements.Therefore,this paper studies infrared and visible images registration methods in complex scenes.The main contents of this paper are as follows:1.We expound the imaging characteristics of infrared and visible images,analyze the registration difficulties of the two kinds of images in complex scenes and introduce the basic image registration methods.2.Aiming at the problem that existing methods often fail to register low-quality infrared and visible images with different resolutions,we propose an infrared and visible images registration algorithm combining rolling guided filter and phase information.Since the existing methods usually use Gaussian filter to build scale-space images to achieve scale invariance,it is easy to lead to image edge weakening and loss of image texture information.In contrast,rolling guided filter has good performance in edge-preserving,so we use rolling guided filter to build scale-space images.Afterwards,in order to extract sufficient salient feature points from low resolution and blurred images,we improve the phase congruency model to locate the structural features of blurred images more accurately,and the salient feature points are extracted from the improved phase congruency maximum moment.For the case that the number of feature points detected in the visible images is much more than that in the infrared images,the invalid features lead to time-consuming.To solve the problem,the common methods delete features based on the strength of the features,which is easy to cause features to gather in the areas with rich texture information,and we use K-means clustering algorithm to delete features of visible images to ensure the features’ uniform spatial distribution.Finally,in order to resist the influence of nonlinear intensity difference between infrared and visible images,we construct the maximum index maps by multi-direction log-Gabor filter response results and phase congruency response values,and feature descriptions are performed on the two index maps respectively.Then they are combined to construct a new feature descriptor.The experimental results show that the proposed algorithm has higher accuracy and universality,and the average registration RMSE of public images is less than 1.5 pixels.In addition,the registration results of engineering application images are better than that of comparison algorithms.3.In engineering application projects,the infrared and visible images’ resolution and clarity are quite different,and the background information is strong while the target information is weak,so there is still an obvious registration error with only one registration stage.We propose a fine image registration method based on infrared saliency maps and 3D feature information.For the weak texture information of the target region,the existing methods can hardly detect features in the target,which directly leads to the background being aligned while the target still has a significant deviation after the registration.Therefore,we extract the visual saliency map of the infrared images and superimpose it on the original infrared image to enhance the target texture information,which lays a foundation for image fine registration.Then,considering the time-consuming and precision requirements in the fine registration,3D feature information descriptors are constructed by the first and second order gradient of images in multiple directions,and SSD similarity measurement of descriptors is performed in the frequency domain by FFT to achieve fine registration.The experimental results show that the proposed algorithm reduces the registration error furtherly,accurate registration is archived of low-quality infrared and visible images with more than 10 times resolution difference,and the registration RMSE of engineering application images is decreased by 0.4 pixels compared with the initial registration results.Finally,an image fusion method based on NSCT transform is used to fuse the image after registration with the proposed algorithm,which verifies the promotion effect of the proposed registration algorithm on the image fusion application.