Representation learning-based star identification method

Patent metadata

Field	Value
Publication number	US-11288830-B2
Application number	US-202015929255-A
Country	US
Kind code	B2
Filing date	Apr 8, 2020
Priority date	May 28, 2019
Publication date	Mar 29, 2022
Grant date	Mar 29, 2022

Abstract

Disclosed in the present disclosure is a representation learning-based star identification method, which utilizes an end-to-end and representation learning based neural network model RPNet, for fast, efficient, and robust full-sky star identification tasks. The RPNet learns a unique star pattern for each star from a huge amount of random simulated star image samples, then a classification is made on these star patterns learned before. The RPNet comprises two parts: (1) a star pattern generator (SPG) based on a star pattern generation network to generate unique star patterns for the star image samples; (2) a star pattern classifier (SPC) to classify the unique star patterns generated on the front end. And a weight search verification algorithm is also proposed in the invention for filtering and verification of main stars of the GRSs identified by the RPNet, which further improves tremendously the identification ability for a single star image.

First claim

The invention claimed is:

1. A representation learning-based star identification method performed by a computing device, comprising:
screening an original star catalog to obtain guiding reference stars (GRSs) available for identification by a star identification network (RPNet);
constructing an initial input for the RPNet according to the GRS and its guiding neighbor stars (GNSs);
constructing an end-to-end and representation learning based neural network model RPNet for fast, efficient, and robust full-sky star identification tasks, wherein the RPNet learns a unique star pattern for each star from a huge amount of random simulated star image samples, then a classification is made on these star patterns learned before, wherein the RPNet comprises two parts: a star pattern generator (SPG) based on a star pattern generation network to generate unique star patterns for the star image samples; and a star pattern classifier (SPC) to classify the unique star patterns generated on the front end;
utilizing a weight searching verification algorithm for filtering and verification of the GRSs identified by the RPNet so as to improve identification rate for the GNSs; and
inputting real star images into the RPNet, and outputting identification results by the RPNet,
wherein in the weight searching verification algorithm, based on characteristics of RPNet outputting probability and different distances between the main stars and an image center, for each star image, three main stars nearest to the image center are chosen to form the initial input of the RPNet for identification; for each chosen main star, the RPNet outputs the probability that the chosen main star belongs to each star in an nSAO star catalog, then chooses three of the most likely stars as candidate stars for the GRS; a position weight and a probability weight are assigned to each candidate star, weight scores of candidate star pairs in a field of view are calculated according to the position weights and the probability weights of the candidate stars, and the GRS with higher confidence ratio is screened out, for star identification, by choosing a candidate star pair with a highest weight score, wherein the nSAO star catalog is obtained by a screening process performed on a Smithsonian Astrophysical Observatory (SAO) star catalog, and during the screening process, a limiting magnitude that can be detected by a star sensor is set to 6.0 Mv, stars in dense areas are deleted, one of double stars whose angular distance between each other is too close is retained, and only right ascension and declination and brightness of GRSs are kept.

2. The representation learning-based star identification method according to claim 1 , wherein the procedure for constructing the initial input for the RPNet comprises: based on distributions between the GRS and its GNSs as well as between the GNSs, choosing a number of m GNSs nearest to the GRS within a pattern radius, calculating angular distances between the GRS and the m GNSs and pair-wise angular distances of the m GNSs respectively, and concatenating the two groups of angular distances subjected to discretization to form a one-dimensional vector as the initial input of the RPNet.

3. The representation learning-based star identification method according to claim 1 , wherein the RPNet comprises the SPG and the SPC, wherein the SPG includes an encoder part of the pattern generation network, and the SPC includes a typical feed-forward neural network; and the pattern generation network is adapted from a denoising auto encoder by adding noise into clean simulated star images so as to construct the discretized noise input; adding a further full connection layer into a traditional encoder-decoder structure that has only one full connection layer, so that the pattern generation network forms a two-layer encoding-decoding structure; and using a sparse encoding strategy to map the initial input into a high-dimensional feature space.

4. The representation learning-based star identification method according to claim 1 , wherein training of the pattern generation network and the pattern classification network is divided into two stages, after finishing the training of the pattern generation network, the pattern classification network is trained according to star patterns and pattern labels generated by the trained pattern generation network.

Assignees

• Univ Beihang

Inventors

• Jiang Jie
• Xu Likai

Classifications

• G01C21/02 Primary

  by astronomical means (G01C21/24, G01C21/26 take precedence) · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V20/13

  Satellite images · CPC title

• G06V10/82

  using neural networks · CPC title

• G06T7/70 Primary

  Determining position or orientation of objects or cameras (camera calibration G06T7/80) · CPC title