Research On Image Matching Method Based On Local Features In Complex Scenes

Nowadays,with the development of science and technology and the continuous updating of image matching technology,the matching method based on image features has developed rapidly,but there are still many deficiencies in practical applications.For example,under complex scene conditions,the image to be matched may contain a large number of repetitive patterns or similar structures,which results in a large number of similar feature extraction,resulting in a reduction in matching accuracy or changes in imaging conditions due to natural factors(such as weather problems)and human factors.Similar scene interference,occlusion and other changes cause the image mismatch problem.The image feature matching system under complex scene conditions faces challenges such as changes in imaging conditions(changes in imaging position,lighting,and shadows),repetitive structures,etc.,which greatly affects the performance of image matching algorithms.This thesis presents my research innovations as follows:1.Aiming at the problem of high mismatch rate of complex scene features containing repeated structures and low efficiency of reliable projection model estimation algorithm,the RANSAC method of estimating projection model is analyzed and improved,and a multi-region reliable random sampling consistency image matching method is proposed.before the image is correctly matched,the similarity neighborhood constraint algorithm is used to achieve stable and reliable feature screening and classification;on this basis,samples are taken from the reliable feature matching set by region,and the initial total feature matching set is used to verify the estimated quality of the model.Experimental results show that this method increases the estimation efficiency of the projection model while increasing the number of correct matching points and improving the matching quality of the image.2.Aiming at the problem of repetitive structure in complex scenes or the presence of some complex image transformations leading to low matching accuracy,the estimated transformation model is used to eliminate the model degradation problems of RANSAC and related methods that mismatch,and the elimination of error matching is converted into a binary classification problem,and an image matching method based on similar learning of corresponding feature neighborhood structure is proposed.This method uses supervised learning support vector machine SVM algorithm,before the real-time image feature matching,the image in the training set is built with the feature points of the neighborhood structure relationship information.The linear classifier is efficiently trained by the combined sequence minimum optimization SMO algorithm.The experimental results show that given the corresponding relationship between the features of image pair with different transformation types,the trained SVM classifier can more accurately complete the classification of correct or incorrect feature pairs,which can effectively improve the accuracy and recall of image matching.