Research On Technologies Of Matching Between Mid-infrared And Visible Band Remote Sensing Images

Image matching is a basic proposition of image engineering, also a fundamental and key technology in the field of computer vision, medical image processing, photogrammetry and remote sensing, etc. The main challenges which faced are from two directions, geometric and radiation distortions between images, as a result the corresponding difficulties of image matching research are wide baseline image matching and multi-modal image matching. Midwave infrared and visible band remote sensing images matching belongs to multi-modal image matching, which has wide application and requirement in remote sensing, computer vision, military and other fields, can provide support for image registration and information fusion of the next stage of image engineering. The projective background of this paper is to make some research on the first domestic airborne mid-wave infrared imaging system. By matching the remote sensing images between the two wavebands, the result of visible band photogrammetric trangulation can be used for the orientation of midwave infrared images. So, there are not only widely applying requirements on development of new airborne remote sensing platform and research of scene matching precision guidance technology, but also significance of promotion on the development of image engineering and the enrich of multi-spectral remote sensing of the earth observation theory, to carry out researchs on matching between mid-infrared and visible band remote sensing images.In view of the current problems of area-based and feature-based multimodal image matching methods, and also the analysis of the characteristics of MWIR and visible band remote sensing images, a fast matching algorithm which is combined with gradient square transform and gradient correlation algorithm, is proposed in this paper. The gradient information outstanding response to the structure characteristics of the image, so the matching which based on grdient correlation in the complex field is more robust and more accurate than that based on intensity correlation in the real field. Thus, gradient correlation matching is further promoted to the fast multmodal image matching. The research work and innovation is mainly reflected in the following aspects:1) Pretreatment technology of matching and characteristic difference between the MWIR and visble band remote sensing images are studied..The diversity and relevance of intensity and geometric characteristics between MWIR and visible band remote sensing images, and alse intensity and geometric pretreatment approachs, which include processing of display enhencement and removal of badpixels of MWIR images, geometric calibration of MWIR camera, and projective rectify treatment, are analysed from principles of electromagnetic radiation and optical imaging.2) The feasible issues of feature-based multimodal image matching are studied.Feasibility premise of feature-based image matching is that the repeatability of feature detection is at the requaired level, then if the repeatability is very low or even to zero, matching will be impossible to discuss. Due to the difference of intensity characteristics of multimodal images, the problem feature missing or nonuniformity of position response is inevitable by using existing algorithms for extracting point or edge features, but the problem is ignored and no effective finding is ever appeared. The bottleneck of feature-based multimodal image matching is not the invariant description of features, but the problem of reproducible feature detection. Thus, The feasibility of feature-based multimodal image matching is very low so far.3) The feasible issues of area-based multimodal image matching are studied.Mutual information is a classical method of area-based multimodal image matching. By analysis and comparison of the main technologies of realization and improvement of mutual information, a simple and practical method to estimate the intensity probability function achieved by smoothing the joint histgrom is proposed. Non-convex and local minima problems are overcomed by combination of PV interpolation technique, thus the multimodal image matching is more feasible.While it is still unfeasible for remote sensing image matching by combination of image structure information with mutual information method for the increased computation complexity, although the the precision and reliability are improved. In this paper, a rapid algorithm with the FFT technique based on self-similarity measure is implemented, and the association between gradient and self-similarity structure imformation is analysised.The research on frequency domain algorithms of gradient-based multimodal image matching is more feasible and useful for the sake of less information loss and lower computational complexity by the representation of complex domain gradient representation.4) An amplitude invariant squared gradient transform algorithm is proposed.With the algorithm, Multimodal images can be of the same gradient orientation properties as the monomodal ones, and the previous problems of singularity and numerical deduction brough with mirrored gradient are overcomed. Compared with the gradient magnitude normalization method, the computational efficiency of the method in this paper is higher, without the trouble of setting the threshold volue, reduce the information loss by retaining the gradient magnitude information, easy to implements fast algorithm, and also can be combined with feature matching methods.5)The rapid algorithm of frequency domain image correlation matching is generalized to the multimodal image matching.Contacts and differences between the frequency domain fast algorithm of intensity correlation and frequence domain phase corrlation are distinguished and analysised for both real intensity images and also complex gradient images. The Fourier-Mellin algorithm is also generalized to the similarty transform matching between multimodal images.Thus, a technology solution of matching between MWIR and visible band remote sensing images is proposed, combined with squared gradient transform、FFT gradient correlation and gradient phase correlation.