Research Of Visible And Infrared Image Matching

Visible images and infrared images are multimodal images.Visible and infrared image matching is an important research area in the field of image matching.Since the visible images can reflect the brightness change and reflection information of the scene with high quality,the infrared image can record the temperature distribution and the radiation information of the scene all day.When the two kinds of image information are fused together,which can provide complemented information and help people enhance the perception and understanding of the target scene.Therefore,visible and infrared image matching,as a key step in image information fusion,have received extensive attention in the fields of military intelligence acquisition,navigation and guidance,remote sensing image fusion,video surveillance and target tracking.However,compared with the traditional single-modal image matching,the difficulty of matching visible and infrared images is that their imaging mechanisms are different,resulting in large gray scale differences between images,and the gradient information of the images cannot provide stable features.Classic gradient-based image matching algorithms cannot be directly used in visible and infrared image matching tasks.Moreover,the more challenging problem is that when there exists geometric variation such as rotation between visible and infrared images,how to construct the similar relationship between the two images and obtain the rotation angle are the keys to the successful application of visible and infrared image matching algorithms.In order to solve the above problems,this paper focuses on the research of visible and infrared image matching methods.In this paper,the characteristics difference and matching difficulty of visible and infrared images are analyzed firstly.Then we introduce the structure similarity which is shared between visible images and infrared images.Based on structure similarity,we proposed a novel invariant feature descriptor which can used in the matching of visible and infrared images,this is the main innovation of this paper.The proposed descriptor employs multi-orientation and multi-scale Log-Gabor filters to encode the edge information statistically.Furthermore,the descriptor provides rotation invariance by estimating the dominant orientation which is based on accumulated edge orientations.The experimental results demonstrate the effectiveness of the proposed rotation invariant descriptor for matching visible and long wave infrared images as compared with state-of-the-art descriptors.We also work on the visible and infrared image patches matching based on deep learning in this paper.We construct six different convolutional neural networks which are based on Siamese network,and proposed different loss functions.What's more,we construct three visible and infrared image datasets to evaluate the performance of different network frames.The experiment results are carefully analyzed and we compare the strength and weakness of the networks.We also propose some methods which can improve the performance of the visible and infrared image patches matching based on convolutional neural networks.In the end,this paper summarizes the research work and proposes the next research plan and outlook.