Fast and dense image matching method and system

Abstract

The invention relates to a fast and dense image matching method and system. The method is characterized in that a matched image is resampled through the local matching of a plurality of specified search regions and through a matching result obtained under a low resolution, so that the search range of a subsequent higher-resolution matching process can be limited through a sliding window, fast and dense global matching between high-resolution images is realized, and meanwhile, the time and space complexity in the matching process is also reduced; in addition, by introducing a matching result based on coordinate expression, topological information of the images is reserved while a matching relation is established, so that the matching process has more continuous and reliable physical significance, and regularization and refinement can be performed on the matching process by utilizing prior information such as parameterized coordinate transformation; and in addition, trainable parameters can be added in the method to realize self-adaptive optimization and be expanded into a universal model.

Description

technical field [0001] The invention relates to the technical field of image data collection, in particular to a fast and dense image matching method and system. Background technique [0002] In recent years, the level of science and technology has been increasing, forming a global automation pattern, followed by the vigorous development of artificial intelligence technology, the main purpose of which is to make machines and computers perceive, understand and act like humans. As one of the most important perception technologies, visual perception plays a pivotal role in this upsurge of artificial intelligence, thus promoting the rapid development of computer vision technology. At the same time, how to understand the differences and connections between multiple visual objects, and how to process the perceived information according to specific needs has become one of the research hotspots in the entire field of computer vision, and image matching is one of the foundations and ...

AI Technical Summary

Problems solved by technology

However, this method needs to repeatedly calculate the gradient of the displacement vector field of multiple scales, and perform iterative optimization to obtain the final result, which still requires a relatively large computational cost

[0008] Patent CN201910580849.3 adopts a global probe-based multi-resolution matching method based on the overall consistency. The corresponding area of ​​the pixel probe after low-resolution matching is used as the limit range of the next layer of high-resolution matching, and the integrated However, due to its characteristics: (a) the search area in high-resolution matching is limited directly by the low-resolution pixel matching results, (b) the global Probes express deterministic matching results in lower-resolution matching, and there are a series of problems: 1) Probes can transmit matching information at different resolution levels at low cost, but they also ignore the low-resolution pixels in the probe area. The possible existence of high-resolution multiple pixel points matching multiple different probe regions limits the density of matching feature points obtained by the patented method in the original image, making it impossible to obtain dense matching results; 2) The final matching result is affected by Low-resolution probe pixels correspond to the position and size of the original image area, so the matching result may depend on the absolute position in the image; 3) In addition, feature (b) makes the mathematical expression of the method non-derivable and cannot propagate gradients , so that it is impossible to add trainable parameter combination machine learning or deep learning algorithms to achieve adaptive optimization or generalization model training for the matching process including feature extraction and screening

Images