Research On Multisensor And Multiresolution-based Image Fusion Algorithms

Image fusion technology is the focus of research in the field of information fusion.It can make full use different redundancy and complementary information obtained bythe same target multiple sensors to obtain more comprehensive, accurate and objectivecharacteristics, widely used in the medical and defense fields and has been widelydeveloped. The first problem to solve of image fusion is image registration, since theimage information of the same object from different sensors in different conditionsmay exist translation, scaling, rotation and other geometric transformation relations, itshould first transformed the images into a unified coordinate system by imageregistration. As registration quality directly affects the fusion effect, thus the imageregistration is an important research topic. The main works of this dissertation are asfollows:(1) The study of theory and methods used in image registration. According to thecharacteristics of infrared and visible image, by analyzing the existing featureextraction algorithm, decided to use SIFT algorithm to extract feature points andestablish feature vectors. Using appropriate distance threshold to select matchingpoints between the extracted feature points.(2) As SIFT algorithm can extract very large quantity points in some occasions,this will lead to great computation in the subsequent steps such as the establishmentand matching of feature descriptor, which has a certain gap with the real-time nature ofthe subject. Therefore, this paper proposes two methods to improve: Firstly, afterdetecting large number of key points in multi-scale space, take them as the initialfeature points, then use the Harris corner detection operator filter these feature points,take the points which have larger corner response function value as the final point ofthe feature point for subsequent calculation. Secondly, when generating featuredescriptor through SIFT algorithm, using principal component analysis (PCA)algorithm, successfully reduced the128-dimensional feature point descriptor generatedby the original algorithm to36dimensions, greatly reducing the amount ofcomputation. Experimental results show that these improved methods greatly improvedthe speed of the algorithm without affecting the accuracy of the registration situation,which meet real-time requirements. (3) Using RANSAC algorithm to filter out the preliminary feature pointsgenerated by SIFT, remove the mismatching feature points, and then solve thegeometric transformation matrix through the least-squares algorithm, using thetransformation matrix to convert the image which is to be addressed.(4) Finally, this paper studies the fusion algorithm of infrared and visible images.First of all, introduced the common fusion algorithm, such as the weighted averagefusion algorithm, Laplace pyramid image fusion and the wavelet fusion algorithm. Forthe lack of discrete wavelet (DWT) transform, proposed the use of image translationinvariant wavelet transform (SIDWT) fusion algorithm based on integration. Forimages gained from different angle of view, this article deals different parts withcorresponding way. Through the objective evaluations, we can see the method greatlyimproves the fusion effect.