SAR target recognition method based on transfer learning and full connection layer output

Abstract

The invention relates to an SAR image target recognition method based on transfer learning and full connection layer output. The SAR image target recognition method comprises the steps: reading SAR image data, carrying out the target interception, and enabling to-be-recognized target images to be consistent in size; carrying out four kinds of preprocessing on the intercepted image data, includingcutting, rotating, zooming and mirror image overturning transformation, so as to expand a data set; normalizing the numerical value of the expanded data set; designing a full-connection layer full-output deep learning model based on a ResNet50 network; importing ResNet50 network parameters trained by utilizing the optical image data set into a design model, freezing bottom-layer parameters of themodel, and training unfrozen parameters in the model by utilizing labeled training samples of the SAR image data set; and predicting the SAR image to be identified by using the trained model to obtainan identification result.

Description

technical field [0001] The invention relates to SAR remote sensing image processing technology, in particular to the field of SAR remote sensing image ground vehicle target recognition technology. Background technique [0002] Automatic target recognition (ATR) of SAR remote sensing images, especially the recognition of ground tactical vehicle targets, has high military application value. The classic target recognition method needs to manually design image features and classifiers, and its performance is limited. It is gradually replaced by the SAR target recognition method based on deep learning. Deep learning designs a neural network that includes multiple convolutional, pooling, and fully connected layers. Under the premise of training with a large amount of data, it automatically extracts target features and completes classification tasks at the same time. The more layers of the network, the stronger the ability to describe the target features. Therefore, it is necessa...

AI Technical Summary

Problems solved by technology

[0003] The SAR image target recognition technology based on deep learning faces three problems: (1) Unlike the massive samples of natural scenes, SAR image ground vehicle targets are difficult to obtain samples and the sample size of the data set is small, which is not conducive to the training of deep networks

(2) Designing a new deep network structure and training it directly from scratch will often lead to slower network iterations due to the large gap between the initial value and the optimal value, and it is easy to fall into local extreme values, resulting in overfitting

This processing method will lose the extracted high-dimensional feature information due to the gradual compression of the training features in the fully connected layer, which will affect the generalization ability of the model.

Images