Research On Dataset Construction And Algorithm For Mobile Phone Image Registration Based On Optical Flow

With the development of smart phones,image multi-frame fusion technology is increasingly widely used in smart phone image processing such as HDR,multi-frame noise reduction,and image super-resolution.Mobile phone image registration is an in-dispensable module of the mobile phone image multi-frame fusion algorithm,and the accuracy of the image registration also directly affects the effect of the multi-frame fusion algorithm.Therefore,it is of great significance to carry out targeted research on the construction of mobile phone image registration data sets and algorithms.The images captured by mobile phones are mostly natural scenes.In the design of the image registration algorithm,factors such as possible foreground motion and multi-plane scenes must be considered.At the same time,the real-time requirements of the registration algorithm are relatively high.Therefore,the research on mobile phone im-age registration technology must be combined with the common shooting scenes of smart phones to improve the registration accuracy while reasonably controlling over-heads such as running time and memory space.Traditional image global registration algorithms are limited in principle and cannot accurately register multi-plane scenes or moving objects;local registration methods based on dense optical flow are not re-stricted by a single transformation matrix and can obtain better registration As a result.the computational overhead is large and it is difficult to realize real-time estimation.In recent years,the dense optical flow algorithm based on deep learning has achieved bet-ter results than traditional optical flow algorithms on the public test set,and has good real-time performance.On this basis,this paper proposes an optical flow algorithm based on deep learning to predict dense optical flow and use it for registration.At the same time,it also proposes a self-built optical flow data set for network supervised learning to improve the performance of the algorithmThe paper first introduces the background and significance of image registration,the current status of domestic and foreign research related to registration,and summa-rizes traditional image registration algorithms based on region information and feature information,image registration algorithms based on optical flow,and light based on deep learning.Streaming algorithms,etc.,through experiments,compare the ad-vantages and disadvantages of the most commonly used traditional image registration algorithms,such as block matching and feature-based image registration algorithmsThe paper introduces optical flow for image registration.Traditional image regis-tration algorithms have the disadvantages of too long registration time or unable to han-dle moving objects due to global transformation.This paper proposes a registration al-gorithm based on optical flow to solve these problems.The paper first introduces opti-cal flow calculation methods,including algorithm principles and optical flow visuali-zation,etc.;then,it proposes an algorithm for registration using optical flow,including image mapping using optical flow,and removing holes and ghosts in the optical flow mapping process;Finally,the optical flow registration algorithm is compared with the traditional algorithm.The experimental results show that the proposed algorithm can effectively improve the registration effect.The paper proposes an improved optical flow registration algorithm and a data set construction method based on deep learning.First select the appropriate network struc-ture,take the appropriate optimization method,and then construct an optical flow data set containing various scenes common in mobile phone images for training;finally,this paper conducts a lot of experiments,and the experimental results show that the pro-posed method can effectively improve the image Accuracy of registration.The data set FlyingThings2D proposed and constructed in this paper has significantly improved the optical flow accuracy and robustness of the optical flow network prediction,making the registration algorithm based on deep learning to predict the optical flow compared to the traditional gradient-based optical flow registration algorithm.Less overhead and higher accuracy..