SAR image target classification method based on multi-kernel scale convolutional neural network

Abstract

The invention discloses an SAR image target classification method based on a multi-kernel scale convolutional neural network. The method comprises the following steps: 1, selecting different types of SAR images as a sample set; 2, carrying out convolution on the input SAR image in parallel by adopting a multi-scale convolution kernel in each convolution layer, and carrying out multi-scale optimization fusion on extracted multi-kernel scale features to obtain fusion features; 3, carrying out the multi-level optimization fusion of the fusion features extracted from the shallow, middle and deep convolution layers, and obtaining the final features; 4, inputting the final features into a full connection layer and a softmax classifier to obtain a prediction result, and comparing the prediction result with a real result to complete a network training process; and 5, inputting the SAR image to be classified into the trained multi-kernel scale convolutional neural network to obtain a corresponding category. According to the method, the target feature representation integrity of the SAR image can be improved, higher classification precision and classification efficiency are obtained, and the method has better engineering application value.

Description

technical field [0001] The invention relates to the technical field of SAR image target classification, in particular to a SAR image target classification method based on a multi-core scale convolutional neural network. Background technique [0002] Synthetic Aperture Radar (SAR) is an active imaging sensor with all-day and all-weather working capabilities. It can effectively identify camouflage and penetrate cover, and obtain high-resolution remote sensing images. Compared with optical images, SAR images can provide a variety of useful information such as amplitude information, phase information and polarization information, which are effectively used in the application of SAR image target classification in the field of military reconnaissance. [0003] Traditional SAR image target classification methods generally obtain the features of SAR images by manual extraction, and then input the features into the classifier for classification. This method is not only time-consuming...

AI Technical Summary

Problems solved by technology

[0004] The existing improved convolutional neural networks basically use a fixed-scale convolution kernel for convolution operations. When the convolution kernel scale is large, the local detail features of the target in the SAR image are easily lost; when the convolution kernel scale When it is small, the overall contour features of the target in the SAR image are easily lost

In summary, the local detail feature extraction and overall contour feature extraction of the SAR image target cannot be taken into account when using a fixed-scale convolution kernel, and the integrity of the SAR image target feature representation is low, making the final classification performance poor.

Images