Picture processing method based on feature point operator of neural network learning basic graph

Abstract

The invention discloses an image processing method of a feature point operator based on a neural network learning basic graph, which comprises the following steps: rendering a data set of a compositegeometric model, and marking points of interest; building a layer normalization full convolutional neural network comprising double branches; using the geometric model data set for neural network training; carrying out adaptive domain migration on the real picture matching set by the neural network to obtain a point-of-interest label of the real picture; obtaining a point-of-interest label of thereal picture; feeding the real picture and the twinning picture of the real picture after Affine transformation into the newest trained neural network in the step S6 for training, and obtaining a probability graph of points of interest and each point of interest as descriptor codes; and processing any two pictures at different visual angles by using a neural network to obtain an interest point anda descriptor thereof of each picture, and matching the feature points of the two pictures. According to the invention, the extraction of the feature point operator of the image is more stable and reliable.

Description

Technical field [0001] The present invention relates to the technical field of camera picture processing, in particular to a picture processing method based on a neural network learning feature point operator of basic graphics. Background technique [0002] The feature points of the picture are the points that have the only representation on the picture under the ideal state. For example, the end points of the line segment and the corner points of the geometric pattern, etc., all have points with more special meaning than ordinary pixels, which is what we need to find Points of interest, in traditional feature point descriptor algorithms, such as SIFT, AKAZE, ORB and other feature points, are often artificially designed pattern extraction algorithms based on the gradient direction, scale invariance, and orientation invariance of the image block to obtain the sense. Points of interest, as well as calculating the descriptors of points of interest, and this method based on tradition...

AI Technical Summary

Problems solved by technology

[0002] The feature points of the picture are the points with unique representation on the picture under the ideal state. For example, the end points of the line segment and the corner points of the geometric pattern, etc., all have points with more special meaning than ordinary pixel points, which is what we need to find Points of interest, in the traditional feature point descriptor algorithms, such as SIFT, AKAZE, ORB and other feature points, are often obtained according to artificially designed pattern extraction algorithms such as gradient direction, scale invariance, and direction invariance of image blocks. Points of interest, and descriptors for calculating points of interest, and this method based on traditional computer image processing often has certain limitations, because it is impossible for people to consider all situations, especially in areas where textures are missing or textures are repeated. , the feature point descriptors are often not stable enough, and the problem of wrong matching often occurs

Images