Research On Panoramic Image Stitching Algorithm Based On Improved AKAZE And RANSAC

Panoramic image stitching technology is a technology that stitches multiple images with overlapping parts into a large-angle image or panoramic image.As an important field in image stitching technology,it has a wide range of military,medical,educational and tourism.As the application scenes of image stitching technology become more and more complicated,a panoramic image stitching algorithm with wide adaptability and high stitching accuracy is urgently needed.However,due to the fact that the feature vectors in the current image stitching technology are often designed manually,the registration accuracy is low,and the scope of application is small,which is difficult to meet people's needs for multi-scenario applications.Therefore,the thesis proposes a panoramic image stitching algorithm based on improved AKAZE(Accelerated-KAZE)and RANSAC(Random Sampling Consensus).The main research contents are as follows:(1)Aiming at the problems of weak stability and low accuracy of manual descriptors used in the image registration stage of traditional image stitching algorithms,a more robust general CNN(Convolutional Neural Network)descriptor was developed instead of manual description Image registration algorithm.First,design the network model and use the GL3D(Geometric Learning for 3D Reconstruction)data set to train to generate the CNN descriptor model.Then use the AKAZE algorithm to construct the non-linear scale space of the image sequence and extract the image feature points,and then use the trained CNN descriptor model and the extracted feature point image blocks to perform a convolution operation to generate the feature point description vector.Finally,the RANSAC algorithm is used to eliminate the wrong matching point pairs and calculate the registration accuracy.The results show that the improved AKAZE algorithm still has a higher registration accuracy when the image viewing angle difference and illumination difference are large,and in various common image changes,its registration accuracy is higher than the traditional AKAZE algorithm.(2)When the feature points are too concentrated,the parameters of the image transformation matrix model calculated are too local,the accuracy of image stitching is low,and the traditional RANSAC algorithm runs inefficiently,so the traditional RANSAC algorithm is improved.The image feature point area is first meshed,and then the image feature points are reduced by only retaining the feature point pairs with the highest matching score in each grid.At the same time,the method of setting the upper limit of iteration number and nesting threshold in the iteration process eliminates the redundant time of matrix model parameter calculation.The results show that the image stitching accuracy of the improved RANSAC algorithm is higher than the traditional RANSAC algorithm,and the calculation speed of the transformation matrix model parameters is also faster than the traditional RANSAC algorithm.