Multi-view SAR target recognition method based on expectation maximization algorithm

Abstract

The invention belongs to the technical field of synthetic aperture radar target recognition, and particularly relates to a multi-view SAR target recognition method based on an expectation maximizationalgorithm. According to the method, the convolutional neural network is used as a feature extractor, the number of the tags is reasonably selected according to the prediction probability of the convolutional neural network to the target, the tag with the maximum single-view prediction probability is no longer simply selected as the prediction category, and the reliable tag is flexibly selected according to the probability distribution of the prediction category. In consideration of information compensation characteristics among multiple perspectives, multiple perspectives of the same target are randomly selected, a multi-perspective label set is constructed adaptively, probability distribution of prediction labels is solved by using an EM algorithm, more accurate estimation labels are obtained, multi-perspective fusion is realized, and the recognition rate is improved. Because the visual angle interval among multiple visual angles is not limited, the method can be flexibly applied toan actual scene, and especially target recognition is carried out on a non-cooperative target.

Description

technical field [0001] The invention belongs to the technical field of synthetic aperture radar (Synthetic Aperture Radar, SAR) target recognition, and in particular relates to a multi-view SAR target recognition method based on an expectation maximization algorithm. Background technique [0002] Synthetic aperture radar is a high-resolution microwave imaging radar that works all-weather and all-weather. It is widely used in military and civilian fields such as battlefield perception and reconnaissance, terrain exploration, and environmental monitoring. SAR target recognition technology is based on the theoretical basis of machine learning and signal processing. It extracts target features and classifies targets. It is a key link in the application of synthetic aperture radar and has become one of the research hotspots in the field of SAR. [0003] Literature [1] constructed a single-input convolutional neural network to solve the recognition problem of 10 types of SAR targe...

AI Technical Summary

Problems solved by technology

[0003] Literature [1] constructed a single-input convolutional neural network to solve the recognition problem of 10 types of SAR target images, but the single-view network did not consider the information compensation and intrinsic complementarity between multiple views, and the classifier directly selects the highest prediction probability The label of is the predicted label, ignoring the fact that the predicted probability corresponding to the real label may be lower than the maximum value, thus causing recognition errors

[0004] Literature [2] proposes a data fusion method based on sparse representation classification (Sparse Representation Classification for fused data, DSRC), explores the inherent complementarity between multiple perspectives, uses multi-view information to solve the problem of SAR target recognition, and obtains better recognition results. However, this method limits the interval between multiple viewpoints. When the interval between viewpoints is greater than a certain range, the recognition accuracy will tend to decline. are often random, so the method is not flexible enough and its applicability is very limited

Images