The basic flow and key technology of image stitching __ image processing

From : mosaic of remote sensing images

The basic process of image mosaic

(1) Image preprocessing: The original image for histogram matching, smoothing filter, enhanced transformation and other digital images
The basic operation of the processing, for the next step of image stitching ready.

(2) Image registration: Image registration is the core of the whole image stitching process, the accuracy of registration determines the quality of the mosaic image. The basic idea is: first, we find the corresponding position of the template or feature point of the image to be matched with the reference image, and then establish the mathematical model of the reference image and the image to be fitted according to the corresponding relation, then convert the pending image into the reference image coordinate system to determine the overlapping region between the two images. The key to accurate registration is to find a data model that can describe two image transformation relationships well.

(3) Image synthesis: To determine the conversion relationship model between two images, that is, overlapping regions, we need to embed the image according to the information of overlapping regions to be mosaic into a visual and feasible panoramic picture. Due to the small differences in topography or the different conditions of shooting, the image gray (or brightness) difference, or the image registration results still exist a certain registration error, in order to minimize the residual deformation or the difference between the brightness (or gray) of the image of the impact of the mosaic results, we need to select the appropriate image synthesis strategy.

At present, domestic and foreign scholars put forward many methods of image registration, but all kinds of methods are related to a certain range of applications and have their own different characteristics. They are generally composed of four elements [9]: Feature space, similarity measurement, search space and search strategy.

(1) Feature space (feature spaces)
The feature space is the information extracted from the image for registration. The characteristic can be the gray value of the image, also can be the structure characteristic such as boundary, contour, or angle point, high curvature point, or statistic characteristic, tall structure description and syntax description, etc. [9-11].

(2) Similarity metric (similarity metric)
Similarity measurement is the similarity between the characteristics of the metric registration image. Typical similarity measures include gray correlation, correlation coefficient, mutual information and so on. Based on the image feature registration algorithm, common similarity metrics are generally based on various distance functions, such as Euclidean distance, block distance, Hausdorff distance and so on. The characteristic space represents the data of the registration, the similarity measure determines the registration degree, the combination of the two can ignore many aberrations which are not correlated with registration, and highlight the essential structure and characteristics of the image.

(3) Searching space (search spaces)
Image registration is a parameter optimization problem, the search space is the space of all possible transformations, that is to estimate the parameters. The composition and range of the search space are determined by the type and intensity of the image transformation. Image transformation is divided into global transformations and local transformations. The global transformation takes the whole digital image as the research object, a parameter matrix is used to describe the transformation parameters of the whole image, and the common global geometric transformations include affine transformation, projection transformation and non-linear transformation, and the local transformation allows the transformation parameters to have position dependence, that is, the transformation parameters of each cell vary with the position of the unit. Registration algorithm is to find a place in the search space to make the similarity between the image measurement of the best position.

(4) Search strategy (searching strategy)
The search strategy is to find the optimal estimation of transformation parameters such as translation and rotation in the search space by using appropriate methods. This is of great significance for reducing the computational volume of registration features and similarity metrics. The more complex the transformation between images, the more complex the search space, the higher the requirement of search strategy, so it is very important to choose the appropriate exploration strategy. The common search strategies include exhaustive search, heuristic search, generalized Hough transform, Multiscale Search, tree and graph search, sequential decision, linear programming, neural network, genetic algorithm, simulated annealing algorithm and so on. Each method has its advantages and disadvantages, and to a large extent, the choice of search strategy depends on the nature of the search space.

When designing the registration algorithm, we first determine the type of image and the range of imaging distortion, then determine the feature space and search space according to the performance index of the image registration, and finally select the appropriate search strategy to find the best matching relation model which can measure the similarity between the image to be matched. According to the different selection of the above four elements, different classification methods of image registration techniques are produced. At present, the algorithm for image registration can be divided into three types: the matching algorithm based on region gray correlation, the matching algorithm based on feature correlation, and the algorithm based on interpretation similarity [12]. Image Registration Technology has been studied for many years, each kind of method contains different concrete realization ways to meet the specific problems. Among them, the image matching based on interpretation needs to be built on the expert system of picture automatic interpretation, so far, no breakthrough has been made.