Research And Application Of Infrared And Visible Image Registration Algorithm Based On Common Features

With the rapid development of computer science,information technology and other related fields,the popularity of various types of sensors has been greatly improved.At the same time,people's demand for the acquisition of images of different modes and frequency bands is also growing rapidly.The scene information captured by a single sensor is limited to a certain extent and is difficult to meet the needs of more applications.Due to the natural complementarity of the information acquired by different types of sensors,it can be applied in different application scenarios.Multi-source image fusion technology aims to analyze and integrate various images of different frequency bands and different modes so as to obtain more comprehensive imaging information.The most typical task in this field is infrared and visible image fusion,which has been widely used in intelligent monitoring,automatic driving,environmental monitoring and other aspects.As the precondition of image fusion,image registration technology plays a decisive role in the fusion effect,so it has been a research hotspot in academia and industry.In this paper,the traditional algorithm in the infrared and visible image registration task experiments,analysis of the traditional algorithm for the multi-spectral registration task bottleneck: feature descriptor for the abstract description of common features performance.This paper will adopt the idea from simple to complex to reach the final goal gradually.Firstly,The relatively simple single-spectral Aerial Image and satellite Image registration algorithm was experimentally studied.Through self-labeling and training using The Inria Aerial Image Labeling Dataset,feature descriptors with certain robustness for perspective transformation and scale change were obtained.Experiments on University-1652 dataset verify the effectiveness of the proposed algorithm.Due to the circled flight of the aerial image for the ground target,the collected aerial data may have the characteristics of large Angle of view transformation and rotation.The deficiency of the classical algorithm for large Angle of rotation is well solved,and a relatively ideal effect has been achieved.Secondly,for multi-spectral image registration,due to the different imaging principles of sensors between infrared and visible images,the gray value of the same scene presents a nonlinear mapping relationship,which is difficult to achieve the ideal effect of the traditional manual design based feature extraction infrared and visible light registration algorithm.Because the neural network has high abstract description performance for features,we want to obtain a feature descriptor with high abstract description performance for the common features of infrared and visible images through training.Firstly,we used the pre-trained feature point detection network to label the infrared and visible images in the RGB-NIR Scene Recognition Dataset Dataset,so as to obtain the trained labels.Then the infrared and visible images are used as the input to re-train the pre-training network to introduce the common feature information of infrared and visible light.The experimental results in XT2 acquisition of campus scene infrared and optical image registration show that the algorithm obtained after training can achieve ideal results.Finally,we demonstrate the validity of the algorithm by applying it to image interpolation and fusion algorithm.