Research On UAV Image Stitching Based On Fusion Feature Of Reduced Dimensional CNN

With the development of UAV(unmanned aerial vehicle)technology and the continuous updating of airborne aerial equipment,UAV image is widely used in different fields such as traffic flow monitoring,map reorganization,accident analysis etc.A single UAV image has limited viewing angles and covers less information.In practice,it is necessary to stitch multiple UAV images into a larger image,which has higher resolution and no obvious seams.In order to solve the existing problem of noise and geometric distortion of UAV image,A fusion of Convolutional Neural Networks(CNN)for the UAV image mosaic has been studied in this paper.CNN is a part of deep learning theory which is excellent in image recognition and classification.the main research content in this paper includes the following aspects:(1)For the structural redundancy and noise problem caused by the high output dimension of CNN high-level network,A principal component analysis(PCA)method is applied to reduce the dimension.First,the matrix decomposition is performed by Singular Value Decomposition(SVD),and then the high-dimensional data information of the CNN fully connected layer is mapped to the low-dimensional space by PCA.Finally,the euclidean metric is used to verify the similarity of image features before and after dimensionality reduction.Experiments show that the CNN model after proper dimensionality reduction improves the accuracy of image feature representation.(2)For the color information lacking,operation time consuming and mismatching problem of scale-invariant feature transform(SIFT)algorithm,The color invariant is added to the original SIFT algorithm to form the color scale invariant feature transform(CSIFT)algorithm.Also this algorithm purify the registration result by using homography matrix and random sample consensus(RANSAC).The experiments proved that the improved CSIFT algorithm matches the feature point pairs with higher correct rate and better registration effect.(3)To solve the high-level feature expression lacking problem of the CSIFT.An algorithm for image registration combined with the reduced dimension of CNN features and CSIFT features is proposed.The experimental results showed that the image registrationalgorithm of the fusion feature has higher precision of feature point matching,and the average running time of CNN after dimension reduction is shorter than that before.(4)For the misalignment,ghosting and stitching problems in UAV image,the Poisson fusion algorithm is proposed.Compared with the effects of weighted average algorithm and Poisson algorithm these two methods.It proved that the Poisson fusion algorithm is more effective.