(A) study on target recognition for multi-polarization SAR images using deep convolutional neural networks

Abstract

Generally, in the case of SAR images, when single polarization data and multiple polarized data are mixed, each model should be learned respectively in order to recognize a target in a SAR image using data-driven learning based models. This thesis presents a study on automatic object (or target) recognition for SAR data with multiple polarized images as well as single polarized images using a deep convolution neural network, which is one of the data driven learning based models. We have studied the method of adjusting the scale of multi-channel input which makes it possible to simultaneously learn single polarized and multi-polarized SAR image target recognition in one deep convolution neural network. By adjusting the scale of the input data according to the number of polarization inputs, the distribution of the output feature map values is present within a certain range. A network named ATRCNN-I then extracts the coherent feature values to construct the feature map, thereby enabling learning with a high recognition rate. If the input data scale is not adjusted according to the number of multiplexed polarization inputs, many convolutional filters have a very small coefficient value, resulting in a meaningless feature map value output, which makes no meaningful contribution to the final output value of the network. It not only increases the complexity but also makes the network learning difficult. On the other hand, it is confirmed that if the dynamic range of the output feature map value of the first convolution layer is adjusted to have a similar range regardless of the number of polarization inputs by adjusting the dynamic range of the input value according to the number of polarization inputs, training on a network with a high recognition rate is possible because the feature map contains various information. In addition, this paper proposes a feature map scaling network named ATRCNN-II that can improve the recognition performance by normalizing the dynamic range of the feature map values of the convolution layer by adjusting the input scale. The feature map scaling neural network is able to utilize a higher level of correlation between different polarization data, which leads to better performance than that obtained with an input scaling neural network. Feature map scaling networks can recognize a target with a 99.51% rate compared to 99.39% with a network separately trained with 4 polarization data only. Also, the proposed feature map scaling network can recognize a target at rates of 98.05%, 98.42%, 98.66%, and 98.05% compared to 96.71%, 97.69%, 98.90%, and 98.29% with networks separately trained with HH, HV, VH, and VV.