Research On GPU-Based Multi-resolution Infrared And Visible Image Registration

In the process of infrared and visible images registration, mainly as in the same scene the scale ofinfrared images which have low contrast and low spatial resolution are large, and the scale of visibleimages which have high contrast and high spatial resolution are small, the proper registrationalgorithm is complicated and the accuracy is not high.This paper uses the scale-space theory to study the algorithm of infrared and visible imagesregistration, presents the method of using multi-scale feature points and edges of infrared and visibleimages as features in the registration, then puts forward the method of using characteristic scale todetermine the sub-images size for the similarity matching and using the LTS-Hausdorff distance toevaluate the similarity of the sub-images. The algorithm of feature extraction in single scale imagescan only get single features, which is not very robust for different scale images registration. Scalespace theory and multi-scale features extraction algorithm can describe the image morecomprehensive. After getting the match pairs, this paper uses random sample and consensus algorithmto delete the false match pairs to gain the correct parameters of space transformation, then uses theparameters to registrate and merge the infrared and visible images. Comparing with the SIFTalgorithm, the method of this paper can match infrared images and visible images better.If the algorithm uses the CPU to serial compute will take a long time, in order to solve thisproblem, the paper uses GPU based CUDA architecture to achieve some algorithems which wouldwaste more time in the whole algorithm, such as the median filtering algorithm, feature pointextraction algorithm and building scale space algorithm. After dividing the images to many blocksand optimizing the algorithm by the advantage of parallel computing of the GPU, the speed of thealgorithm is faster. The results show that the GPU and CPU speed ratio of some parts of the algorithmexecution can achieve several hundred times.