Image Mosaic Algorithm Based on PCA-ORB Feature Matching

: In the process of image stitching, the ORB (Oriented FAST and Rotated BRIEF) algorithm lacks the characteristics of scale invariance and high mismatch rate. A principal component invariant feature transform (PCA-ORB, Principal Component Analysis-Oriented) is proposed. FAST and Rotated BRIEF) image stitching method. Firstly, the ORB algorithm is used to optimize the feature points to obtain the feature points with uniform distribution. Secondly, the principal component analysis (PCA) method can reduce the dimension of the traditional ORB feature descriptor and reduce the complexity of the feature point descriptor data. Thirdly, KNN (K-Nearest Neighbor) is used, and the k-nearest neighbor algorithm performs roughly matching on the feature points after dimensionality reduction. Then the random matching consistency algorithm (RANSAC, Random Sample Consensus) is used to remove the mismatched points. Finally, the fading and fading fusion algorithm is used to fuse the images. In 8 sets of simulation experiments, the image stitching speed is improved relative to the PCA-SIFT algorithm. The experimental results show that the proposed algorithm improves the image stitching speed under the premise of ensuring the quality of stitching, and can play a role in fast, real-time and large-scale applications, which are conducive to image fusion.

of feature descriptors to obtain a more uniform distribution of feature points, which speeds up the construction of descriptors.
At the same time, KNN algorithm is used to roughly match the feature points after dimension reduction, and then adopt The RANSAC algorithm eliminates the mismatched points, and finally uses the fade-in and fade-out fusion algorithm to fuse the images.

FAST Algorithm: The FAST algorithm was proposed by
Rosten (Rosten et al., 2006) in 2006 and was revised in 2010.
The algorithm is just a feature point detection algorithm, but the feature description processing cannot be implemented. The algorithm determines whether it is a feature point by comparing the pixel gray level around a candidate pixel. The main algorithm is described in the literature (Rosten et al., 2006;Liu et al., 2017). However, the feature points detected by the FAST algorithm have no direction. To compensate for this defect, the FAST feature points need to be added with direction information. Therefore, the ORB algorithm uses the oFAST algorithm. The basic idea: assuming that the intensity of a corner point deviates from its centroid, we represent this deviation distance as a vector defined as the formula (1)  2.1.2 The BRIEF Algorithm: The BRIEF algorithm is a binary string-based descriptor that is simple and fast to calculate. The rBRIEF (rotated BRIEF) algorithm is an improvement of the BRIEF algorithm to solve its rotation invariance. The main algorithm is described in the literature (Zhou et al., 2015;Li et al., 2017).
In the pixel area around 31 31  feature point, select a sub-window of 5 5  and find the sum of the gray values of the sub-window. The expression of the gray value is as follows: Where, p(x) and p(y) are the sum of the gray values of two sub-windows respectively. The BRIEF feature descriptor is recorded as a binary test vector containing n dimensions, forming a descriptor of the feature point. The expression is as follows: Where n is 256. For   , i i x y at any position, n-dimensional binary test results in a n  2 dimensional matrix: The corner point direction is known by the FAST algorithm, and as showed by the formula (3), the rotation matrix is obtained as .
Finally, the descriptor is rBRIEF:

PCA Algorithm
The idea of PCA algorithm is to map n-dimensional features to k-dimensional (k<n), which is a different orthogonal feature.
We call this k-dimensional feature a principal component and a reconstructed k-dimensional feature. Rather than simply removing the remaining n-k dimension features from the n-dimensional features. It reduces feature counts, reduces noise and redundancy, and reduces the likelihood of over fitting. The main algorithm is referenced in the literature (Vinay et al., 2015;Jiang et al., 2016).
With a data matrix of n samples, which is Calculate the covariance matrix S of the above formula: The eigenvalues of the covariance matrix S is obtained, and the eigenvalues are sorted in descending order, and the largest k of them are selected, and then the corresponding k eigenvectors is respectively used as column vectors to form an eigenvector matrix.
The corresponding feature vector is: So, for the sample x in any n-dimensional space, there is a matrix: So, the matrix Z is the result of the dimensional reduction of the matrix X.

Feature Point Rough Matching
After generating the feature descriptors of the two graphs, the two graphs need to be feature point matched. This paper uses the K nearest neighbor algorithm, also known as KNN (K-Nearest-Neighbor) algorithm. The algorithm is a simple algorithm for data classification. The core idea is that most samples of the most neighboring samples in the feature space belong to a certain category, and the sample also belongs to this category. In plain terms, the classification principle of the algorithm is the principle that the minority obeys the majority (Lu et al., 2017). Then use the LSH (Locality-Sensitive Hashing) search algorithm to quickly find the matching point.

Elimination of Mismatch Points
Feature point matching is performed by the feature point matching method described above, and there are still unavoidable mismatch points in the matching result. In order to ensure the accuracy of image matching, it is necessary to use corresponding processing methods to eliminate these mismatch points. In this paper, random sampling consistency (RANSAC) (Zhang et al., 2016) algorithm is used to eliminate the mismatched points. The RANSAC algorithm is a robust algorithm for transformation estimation. It optimizes the parametric model to eliminate the data of the "inside point".
Unreasonable "outside points" to improve the correct rate of

Image Fusion
After the image is registered, the spatial relationship between the images has been determined. At this time, the fusion algorithm has to be selected. In the selection of the fusion The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W10, 2020 International Conference on Geomatics in the Big Data Era (ICGBD), 15-17 November 2019, Guilin, Guangxi, China algorithm, we should pay attention to the following two key elements: one is to carry out the fusion operation only for the overlapping area and try to keep the original image information unchanged; the other is to maintain the smooth transition of the fusion boundary, which can effectively avoid the spelling generated during the splicing (Yang et al., 2002). So, this paper uses the gradual involution fusion algorithm proposed by Szeliski (Szeliski, 1996).
The specific fusion algorithm is as follows:

EXPERIMENTAL RESULTS AND ANALYSIS
In order to verify the feasibility of the proposed algorithm, the PCA-ORB algorithm and the PCA-SIFT algorithm were used to perform the splicing experiment on the 8 groups of images to be stitched. The experimental environment is MatlabR2014a, and the operating environment is Intel(R) Xeon(R) CPU E5645 2.40GHz, and the memory is 4GB 64-bit Windows operating system. Due to space reasons, 8 groups of images to be stitched cannot be displayed. This paper only shows the matching and fusion results of the 3 groups of images. Therefore, Figure 1 shows the images to be stitched, Figure 2 shows the matching result of the PCA-SIFT algorithm, Figure 3 shows   It can be seen from Table 1 that the matching logarithm of the algorithm is reduced compared with the PCA-SIFT algorithm, but the more stable matching pairs are retained, and the The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W10, 2020 International Conference on Geomatics in the Big Data Era (ICGBD), 15-17 November 2019, Guilin, Guangxi, China matching rate is higher. It can be observed in Figure 5 that the image splicing time consumption of the algorithm (PCA-ORB algorithm) is less than that of the PCA-SIFT algorithm.
Comprehensive Table 1 and Figure 5 can be concluded that the PCA-ORB algorithm has great advantages in feature point matching, which is conducive to image stitching.