Research On Image Matching And Correction Based On Scale Invariant Feature Transformation

Image matching technology in computer vision is a fundamental problem in the field of image processing and plays an important role in intelligent driving,pose estimation,image alignment and other fields.Image matching algorithms have been developed for a long time,but there are still some shortcomings in practical applications because of the limitations of traditional algorithms and deep learning itself.For example,the accuracy of the algorithm is not high,and the generated matching point pairs have more mis-matches;the generalization is not strong,and the results are often good only in specific scenes or data sets.In order to explore a more reasonable image matching solution and apply it,this paper conducts an in-depth study and analysis of the scale-invariant feature transform algorithm(SIFT)from the scale-invariant theory,and the main works include:(1)The main methods of image matching are grouped into three categories with respect to different matching objects and research methods,and the principles of each type of algorithm are summarized and elaborated,focusing on feature-based matching,and the core principles of convolutional neural networks(CNN)are analyzed to summarize their advantages and disadvantages.(2)An improved SIFT algorithm is proposed,which improves the traditional SIFT algorithm in three aspects:key point localization,descriptor generation and false match rejection.In the key point localization,the number of key points is reduced by using more information of DoG adjacent layers to constrain the selection of extreme value points;in the descriptor generation,the neighborhood division of key point descriptor generation is changed and the descriptor dimension is reduced by means of concentric circular neighborhood to improve the anti-rotation ability;in the mis-match rejection,the RANSAC process is improved and two sets of clusters are generated by the clustering algorithm The RANSAC process is improved in terms of mismatch rejection,generating two sets of clusters by the clustering algorithm,and selecting the appropriate set for model estimation according to the scoring function,which effectively avoids the problems of blindness of the initial data of the algorithm and too many iterations.The experiments show that the method has more than 96%correct matching rate in four scenes of affine,illumination,rotation and compression.(3)A method combining the improved SIFT and HardNet CNN for feature extraction is proposed.The coordinates,scale and angle information of feature points are first calculated by the improved SIFT algorithm,then the neighborhood size of feature points of corresponding scale is calculated,rotated transformed and downsampled,and input to HardNet to calculate feature descriptors by seven-layer convolutional network.Compared with the improved SIFT algorithm,the descriptors extracted using the convolutional neural network are deeper and reduce the matching error rate of the fuzzy transform by 50.16%while keeping the other scene matching effects stable.(4)Complete the design of image matching and correction system,package and integrate the algorithm into the system through back-end development technology,complete the design of API,verify the feasibility of the algorithm and simplify its use in practical engineering.