基于深度学习的低频SKA带宽涂污效应矫正方法

A Correction Method of Bandwidth Smearing Effect for Low Frequency SKA Using Deep Learning

摘要: 针对低频SKA成像中的带宽涂污问题，本文提出了一种基于深度卷积神经网络的带宽涂污效应矫正方法。所提方法设计了一个带宽涂污效应的矫正模型，该模型通过其中的残差学习机制对低频SKA观测信号中的河外点源信号携带的带宽涂污效应进行特征提取，进而提高了对带宽涂污效应的矫正效果。首先需要对河外点源信号进行带宽涂污效应图像模拟，并对模拟图像进行电磁干扰噪声添加、归一化操作，再利用矫正模型中批归一化与卷积相整合的结构，进而提取图像深层次的噪声特征，最后利用残差学习技术，减轻神经网络负荷的同时完成图像重构。实验结果表明，该深度学习方法可以在强带宽涂污效应下达到良好的矫正效果，同时具有良好的鲁棒性，以及可移植性。

Abstract: In order to solve the problem of bandwidth smearing in low-frequency SKA imaging, a correction method based on deep learning of bandwidth smearing effects is proposed in this paper. The method designs a correction model of bandwidth smearing effect, which uses the residual learning mechanism to extract the characteristics of the bandwidth smearing effect carried by extragalactic point sources in the low-frequency SKA observation signal to improve the correction result of bandwidth smearing effect. Firstly, we need to simulate the bandwidth smearing effect image of the extragalactic point sources in the low-frequency SKA observation, the electromagnetic interference noise is added and normalized to the simulated image, and then use the structure of deep convolution layer and batch normalization layer in the correction model which extracts the deep-level noise features of the image, and finally residual learning technology is used to reduce the learning load of network and reconstruct the image. It is found that the deep learning method has great correction performance with strong bandwidth smearing effect, good robustness and portability.