Fast And Robust Image Stitch Based On Sparse Features

Sparse features matching is the basis of computer vision algorithms.So far,most of the computer vision algorithms are directly or indirectly based on the results of image sparse features matching.In real-time image processing applications,the speed and quality of image sparse point matching are crucial to the processing performance.Aiming at this problem,this paper focuses on a fast matching algorithm of image feature points.The algorithm can quickly match the two images which conform to the homography and achieve the performance of real-time processing.And in the case of poor image quality,this algorithm can still achieve fast and efficient matching.In this paper,the fast matching results are applied to the image stitching application to verify the effectiveness and efficiency of the algorithm.The following are the main contents of this paper:(1)In order to extract image features quickly,this paper uses FAST corner detection algorithm to extract features,and Harris corner response value as the standard to measure the strength of keypoints,then meshes the image and extracts each grid.The feature points with high response values are internalized to achieve uniformity of strong feature points.(2)Binary description is performed for all feature points,wherein the strong feature points are respectively subjected to the two binary descriptions of the BRIEF and ORB,and then the matching is performed separately,and the two matchings are completely matched and filtered out as the candidate matching pairs for calculating the guiding matrix.This allows for quick calculation of the steering matrix between images while providing a priori knowledge of image matching.(3)A location matching method based on guidance matrix is Proposed.Firstly,the target image is meshed,and then the feature point coordinates are used to classify the feature points,and the feature points in the original image are mapped to the position on the target image.The mapped coordinate information can be located in the grid of the target image,and the original image feature points can be violently matched with the feature points in the grid,so that we can achieve fast and robust matching of image features,and improve the matching accuracy and recall rate.(4)To improve the RANSAC,according to the key observation here is that correct hypotheses are tightly clustered together,We decide whether to carry out expensive error projection calculation by judging whether the fixed four points are mapped by two homography matrices and measuring whether the mapped points have conflicts.If there is a conflict,then calculate whether the current homography matrix is the optimal model.If there is no conflict,then return to the random pair to recalculate the new homography matrix.In this way,the filtering of most error homography matrix is completed,and the accelerated operation of RANSAC algorithm is realized.(5)In order to solve the problem of misalignment and ghost problem of the stitched image,the optimal seam and multi-band fusion algorithm are implemented to improve the image quality.In summary,the feature extraction,classification,matching and RANSAC process are improved.The algorithm improves the speed when stitching the image,and the success rate is also greatly improved.The algorithm in this paper has great advantages for solving remote scenes or slightly moving scenes.