Research On Image Matching Method Based On Machine Vision

Image matching is a method of matching similar or identical images by analysing the texture,greyscale and other feature information of the image.It is the most critical step in the research of machine vision and image processing technology,and has been widely used in industrial defect detection,object recognition and localisation,medical image analysis,unmanned vehicles and other fields.Image matching methods are generally divided into two categories,based on grey-scale information and feature information,with the former being computationally complex and time-consuming,and the latter being computationally simple and time-consuming with strong robustness to complex changes.At present,scholars at home and abroad have done a lot of research to improve image matching methods,but in the face of the complex and changing image acquisition environment in practical application scenarios,such as lighting and viewing angles,image matching may still suffer from a decline in accuracy,so it is particularly important to study image matching techniques in complex and changing scenarios.This paper will focus on feature-based image matching methods and propose improved algorithms that can maintain better performance in complex and changing scenes.The main work of the paper includes:(1)The SURF(Speeded Up Robust Features)algorithm and ORB(Oriented FAST and Rotated BRIEF)algorithm,two feature extraction methods based on point features,are highlighted.The basic principles and implementation process of the algorithms are systematically analysed,and the advantages and disadvantages of the algorithms are analysed.Several key techniques of image matching methods,pre-processing methods,basic components,classification of matching methods and evaluation criteria of performance are also analysed to provide theoretical support for the subsequent chapters.(2)Due to the poor rotatability of the traditional SURF algorithm in the face of scenes with large rotational transformations,the feature descriptors are prone to mis-matching after construction.To address this problem,a matching image algorithm with improved SURF feature points is constructed by introducing the DAISY descriptor into the SURF algorithm through an in-depth study of the DAISY descriptor.The algorithm starts with feature detection by the SURF algorithm,then uses the DAISY descriptor to complete the feature description,and finally uses a random sampling consistency algorithm to complete the matching with the initial matching of Euclidean distance and the elimination of mis-matching.In this paper,three sets of images with different changes are used to compare with the traditional SURF algorithm.The experimental results show that the improved SURF algorithm,while ensuring the feature detection capability of the SURF algorithm,reduces the computing time and improves the data processing efficiency compared with the traditional SURF algorithm while maintaining a higher number of correctly matched pairs of points in scenes of rotation,greyscale,image blur,JPEG compression,lighting and other transformations or real images,and also ensures the real-time performance of the algorithm.(3)The ORB algorithm does not possess scale invariance and the traditional image recognition method has low recognition efficiency.Combining the advantages of the performance of the SURF algorithm,an improved ORB-FLANN method for workpiece image recognition is constructed and applied to the workpiece image recognition scenario for verification.The feature descriptors are constructed using SURF for the feature points detected by the ORB algorithm.In the feature matching stage,for the original FLANN matching will have the problem of one-to-many and other mis-matching,a two-way nearest neighbour FLANN matching method is proposed,through which the coarse matching of the improved algorithm is achieved and the feature points with higher reliability are retained.Finally,PROSAC is used to complete the rejection of other pairs of mis-matched points to improve the accuracy of image matching.The experimental results show that the improved algorithm improves the correct matching rate by 2.6%～18.8% and 29.5%～43.9% respectively when processing different transformed images compared with other existing algorithms in the literature,and the relative algorithm time consumption is controlled within 4s,while the recognition efficiency of the artefact images is improved,and the improved algorithm has scale invariance.