Feature Points Matching Based On Convolutional Neural Networks

Multi-mode,multi-angle image registration fusion in the three-dimensional reconstruction,medical image processing and three-dimensional display has great application value.If you can display a number of modal information on an image,it will greatly provide the basis for the doctor to judge the condition,and the key to multi-mode image fusion is image registration.Image registration is the process of finding the best match for multiple images obtained under different conditions(time,equipment,brightness,shooting position and angle,etc),which achieves the purpose of multiple images can be displayed on an image.In other words,image registration is looking for the spatial mapping relationship,and making the two images correspond to the same location in the space one by one.Image registration is divided into three categories: gray,transform domain and feature-based image registration methods.Image registration based on feature with its strong anti-interference ability to noise,deformation,and the smaller computation and higher robustness,is widely used.The SIFT(scale invariant feature transformation)algorithm can be invariant to scale,rotation and illumination.It is the most effective feature points matching algorithm of the feature-based registration algorithm.This method mainly solves the registration problem of homologous image with scale,rotation and illumination transformation,and it is very ineffective for multi-mode image registration.This paper introduces a machine learning method for image registration.The Siamese convolution neural networks is trained by a large number of images with scale,rotation,illumination changes,perspective transformation and non-rigid deformation.The 128-dimensional vector extracted by Siamese convolution neural networks can resist scale,rotation,illumination changes,perspective transformation and non-rigid deformation.Compared with the gradient feature information extracted by the SIFT algorithm,the feature vector extracted by Siamese convolution neural networks is richer,comprehensive and more relevant,which is helpful to the matching of image feature points.Due to the different imaging methods,there is a certain degree of complexity in image registration of multi-mode images.It is not good to get the registration method by the conventional method.Based on the introduction of Siamese convolution neural networks for feature vector extraction,this paper combines the SIFT feature point extraction method for multi-view image processing,and achieves better results for multi-angle image registration,and it can also get a good registration effect for the multi-mode image registration.