Binocular Image Matching Method Based On Convolution Neural Network

With the continuous advancement of machine vision technology,there is a high requirement for the accuracy of target locating.In a multi-target environment,due to the external environment or the image texture is more abundant,the matching accuracy of the target will be reduced.Locating technology in drones can have a big impact.The matching accuracy will directly affect the locating accuracy of the SLAM algorithm based on feature point matching.Feature matching is also the basis of many computer vision problems,such as the recognition in the vision or the motion structure problem,so the matching accuracy of the feature points is a direction of great research value,and it is also a Challenging problems.At present,the main method of removing mis-matching is to rely on a large amount of random sampling consensus algorithm and violence matching algorithm,which makes the calculation complexity is high,and the number of algorithm sampling is not set well.In order to obtain the high-precision matching information of the recognition target in the image,with the great development of deep learning,this paper studies the method of removing mis-matching based on convolutional neural network.Based on the ImageNet database,the convolutional neural network is trained to stimulate the neurons.The computational load is reduced by weight sharing,and the parameters are adjusted by gradient descent.The target is identified based on the area recommendation network,and the target is subjected to pixel-by-pixel convolution processing to obtain target pixel level cutting information.Finally,the matching of the binocular data sets is de-matched and matching identification information of the matching region is obtained.If the tag information is matched,the matching points in the target contour region are reserved.This matching method has better robustness and has been well verified by simulation tests.The major tasks completed include:1.A brief introduction to the research significance of machine learning,driverless and feature point matching,and research progress at home and abroad.2.The matching of several typical feature points is studied,including the algorithmprinciple that the algorithm has scale invariance,direction invariant and anti-noise ability.And analyzed the pros and cons between them.3.The structure,propagation,training,and the core algorithm of the convolutional neural network are carefully studied.In order to improve the accuracy of ORB matching in multi-target images,an image ORB removal mis-matching method based on convolutional neural network is proposed.The algorithm firstly recognizes the image by Faster R-CNN method,and uses the area recommendation network to obtain rectangles.Label the region of interest and the category label,this step can obtain the prediction category and coordinate information of the region of interest,and perform pixel level correction through the full convolutional network convolution layer to obtain the category of the target at the pixel level and then perform target segmentation.Finally,on the basis of the matching of the original ORB feature points,the mismatched points outside the same target segmentation area in the two images are eliminated.5.In order to verify the validity of the method,the simulation experiments were performed on the traditional ORB matching and ORB matching based on this method,and the experimental data was analyzed.