Classification based optimal time-frequency distribution design and target identification method

Abstract

The present invention relates to a classification based optimal time-frequency distribution design and target identification method, which comprises a design process and an identification process; The design process comprises: SA1, calculating a fuzzy function of a training set signal and a fuzzy function mean; SA2, selecting a two-dimensional radial gaussian kernel function as an optimal kernel function; SA3, iteratively searching and calculating the optimal kernel function; SA4, performing time-frequency transformation under the optimal kernel function on the training set signal, and extracting a feature value; SA5, designing a classifier of the training set signal, and classifying the feature value. The identification process comprises: SB1, performing time-frequency transformation under the optimal kernel function on a testing set signal, and extracting a feature value; and SB2, according to the classifier of the training set signal obtained in the design process, performing target classification and identification on the testing set signal. According tot the method provided by the present invention, two separate links, i.e. a feature extraction algorithm and classification design, are combined by means of an optimization process of the optimal kernel function, so that the feature value acquired by the feature extraction algorithm is advantageous for the classifier design, and accuracy of a target identification system is effectively improved.

Description

technical field [0001] The invention relates to a classification-based optimal time-frequency distribution design and target recognition method, belonging to the technical field of target recognition. Background technique [0002] In the field of target recognition, feature extraction algorithm and classifier design are the two most critical links. The quality of the feature extraction algorithm will directly affect the difficulty of classifier design, and then affect the accuracy of the target recognition system. Features can generally be divided into point features, line features, and regional texture features, etc., which are used to describe a collection of target features. The more eigenvalues, the more detailed the description of the target, and usually the higher the accuracy of the target recognition system; but with the increase of eigenvalues, the eigenvalues ​​of the same type of target data (or images) will also differentiate. Increasing the eigenvalue will not...

AI Technical Summary

Problems solved by technology

[0004] In the prior art, among the patents in the field of target recognition, there are about feature extraction: the application number is CN201110257384.1 based on Radon (Radon) transformation and polar harmonic transformation invariant moment target recognition method, the application number is CN201210148117 .5 radar target recognition method based on radar target distance image time-frequency feature extraction, and a target recognition method based on affine invariant moments of key points with application number CN200910076361.3; but these patents are A fixed transformation method is used, so it has certain limitations in target recognition. If several types of targets have basically the same characteristics in the transform domain, the recognition of these types of targets cannot be realized.

In addition, the application number CN201410371180.4 is based on the matching dictionary and compressed sensing radar one-dimensional range image target recognition method, which is to extract the dictionary from the target and use the dictionary information to identify the target, but this method is directly from the signal itself Extracting the feature sequence, rather than through time-frequency transformation, most nonlinear signals do not have obvious feature differences in the time domain or frequency domain, but have certain differences in the time-frequency domain, so they cannot be directly obtained from the time-frequency domain. The feature sequence that can be used for classification can be obtained in the domain or frequency domain, and the feature sequence can only be extracted in the time-frequency domain

Images