Small sample remote sensing image target identification method based on metric learning

Abstract

The invention discloses a small sample remote sensing image target identification method based on metric learning, and belongs to the technical field of remote sensing image identification. The methodcomprises the following steps: searching remote sensing image data, and preparing a remote sensing image target slice data set for small sample target identification; performing feature extraction onthe multi-scale remote sensing image target slice by using a deep residual network to obtain a feature vector of a test target and a feature vector set of various training samples; finding the firstk feature vectors similar to the feature vector of the test target in the feature vector set of each type of training samples based on a k-nearest neighbor algorithm; finally, based on a metric learning method, calculating the similarity between the test target and various training samples, and completing identification of the test target. The method has the characteristics of simple network structure, simple training method, high average target recognition accuracy and the like, and is suitable for recognition application of remote sensing image targets in small sample scenes.

Description

technical field [0001] The invention relates to the technical field of remote sensing image recognition, in particular to a small-sample remote-sensing image target recognition method based on metric learning, which can be used for target recognition of visible light remote-sensing images in small-sample scenes. Background technique [0002] At present, although there are a large number of remote sensing images, in some specific scenarios, for some specific targets, the number of samples available is small, and there is a problem of small samples. In the prior art, there are mainly the following methods for small-sample learning in the field of remote sensing, but there are corresponding defects in performance: [0003] (1) The method of sample augmentation generates similar pseudo-samples based on existing training samples to expand the sample size. This method is the most direct, but in some specific remote sensing scenarios, the number of samples available is very limite...

AI Technical Summary

Problems solved by technology

This method is the most direct, but in some specific remote sensing scenarios, the number of samples available is very limited. Even with sample augmentation, the number of effective pseudo-samples that can be generated is very small, and there is still a problem of small samples.

[0004] (2) Based on the migration learning method, this method uses the existing rich sample pre-training target recognition network, and then fine-tunes the pre-training network based on small samples, but this method will be over-adapted, and only for small sample target recognition accuracy is high , while performance drops in the original object recognition task

However, when the samples are scarce, the inter-class differences are relatively weak, which increases the difficulty of target recognition.

Images