Small sample remote sensing ground object classification method and system based on feature generation network

Abstract

The invention provides a small sample remote sensing ground object classification method and system based on a feature generation network. The method comprises the steps of dividing a remote sensing ground object classification data set into a source data set and a target data set based on ground object categories; training a feature extractor and a generation network on the source data set to complete parameter initialization of the feature extractor; on the basis of the target data set, using the feature extractor and the generation network for feature generation to obtain a generation feature set corresponding to each surface feature type, so that high-quality target data can be effectively expanded, and the over-fitting risk on the target data set is reduced; then, using a feature extractor for carrying out feature extraction on the test ground object image; and performing similarity measurement on the test sample features and the generated feature sets of the various ground objecttypes, giving confidence scores of the test images on the various ground object types, and further determining the ground object classification of the tested images, so the remote sensing ground object classification problem under the small sample condition is solved.

Description

technical field [0001] The invention relates to the field of remote sensing intelligent recognition, in particular to a method and system for classifying small sample remote sensing objects based on a feature generation network. Background technique [0002] The deep learning method has been widely used in the remote sensing object classification task and has repeatedly refreshed the classification accuracy on the public data set. However, with the practical application of the deep learning method in the remote sensing object classification task, a major problem exposed is The severe dependence of model classification accuracy on huge training data. [0003] In the practical application of remote sensing intelligent recognition, the collection and labeling of remote sensing image data requires a certain amount of manpower, material resources and time, which forms a contradiction with the dependence of deep learning technology on massive data, mainly reflected in: [0004] (...

AI Technical Summary

Problems solved by technology

[0004] (1) For specific ground object scenes, such as regional remote sensing imaging under severe imaging weather and remote sensing data of a secret military base, it is not enough to collect thousands of high-quality remote sensing images for the training of deep learning models. practical;

[0005] (2) The training data set prepared before model deployment is often difficult to match the actual needs

Its defect is that it needs to train the deep generation model and the classification model separately, the learning efficiency is low, and the learning efficiency is low on small samples facing various ground object classification tasks, and it lacks the versatility of various ground object targets.

Images