Research On Related Problems Of Feature Point Matching In Image Stitching

Image stitching is a technology that integrates two images with overlapping areas into a large field of view image,the stitched image contains the information of the two images,which solves the problem of small amount of information in a single image.As an important part of computer vision,image stitching technology is widely used in remote sensing images,medical images and intelligent driving.There are three main methods for image stitching,the most important of which is feature-based matching.In feature matching,matching of feature points is the most common.Feature point matching has the advantages of fast speed,flexible handling,and high accuracy,which is the basis of image processing tasks.The matching of feature points in image stitching mainly includes three aspects: feature point detection,feature point information description,and eliminate mismatches.Based on the existing technology,this paper studies the feature point information description and eliminate mismatches,the specific content is as follows:1.At present,the description of feature point information is mainly based on the learning method,This approach requires model training on the neural network.Model training requires a large number of highquality samples,this paper aims at the shortcomings of the small number of samples collected by the current sampling strategy.By calculating the class distance of the input samples and extracting hard negative samples with small distances,the training data is enhanced te ensure the samples.while increasing the quality,the number of samples is increased.2.Traditional neural networks use cross-entropy as the loss function to train the network,ignoring the performance constraints on generating descriptors,and can not guarantee the quality of the descriptors.Here,the loss function is divided into three constraints,and the update direction of the neural network parameters is regulated from the distance relatinship between samples,the distance between sample pairs and the bit distribution of the descriptor.Experiments show that can improve the performance of the descriptor.3.Analyze the operation process of the traditional non-maximum suppression algorithm,Based on the implementation of the traditional algorithm,an adaptive non-maximum suppression algorithm is proposed,By sorting and introducing the labeling method,reducing the amount of calculation.Experiments show that the improved algorithm has higher efficiency.4.In order to achieve the feature point de-error matching,this paper also researches the de-error matching algorithm RANSAC,Aiming at the shortcoming of its iterative times,this paper proposes a de-error matching algorithm based on local similarity.Finally,the error-reduction matching algorithm based on local similarity,RANSAC error-reduction matching algorithm and the combination of the two algorithms are used to perform experiments on the image data set to analyze the comprehensive performance of the algorithm.