基于BiGRU-CNN的宽带电磁图像条带噪声去除方法研究

Research on wideband electromagnetic image striping noiseremoval method based on BiGRU-CNN

电磁探测成像系统能够对电磁干扰源进行大范围、宽频带且快速的定位,系统主要由抛物反射面和多通道超宽频带信号采集系统组成。由于各个通道器件参数受限于制造工艺的影响不可能完全一致,探测不同频率干扰源的响应特性也不相同,导致获得的电磁图像中存在的条带噪声随干扰源的频率变化而呈现出不同的特征,严重地影响定位的精度。构建了双向门控循环单元(BiGRU)-卷积神经网络(CNN)模型,根据实测数据构建数据集作为模型的输入,BiGRU和CNN利用图像相邻行间的强相关性,从过去和未来的输入中广泛收集冗余信息,对条带噪声进行提取并对空间信息进行整合处理,利用数据之间的差值对这个过程进行循环迭代。通过大量的实验对模型进行验证,BiGRU-CNN方法与测试的经典方法相比更优,在垂直梯度能量方面降低了15.2%,在残差非均匀性方面降低了21.9%。

The electromagnetic detection and imaging system enables wide-range,wideband,and fastlocalization of electromagnetic interference sources.The system primarily consists of a parabolic reflector and a multi-channel ultra-wideband signal acquisition system.Due to variations in device parameters across channels caused bymanufacturing processes,it is impossible to achieve complete consistency,resulting in stripe noise in the obtainedelectromagnetic images that significantly affects localization accuracy.A bidirectional gated recurrent unit(BiGRU)-convolutional neural network(CNN)model was constructed,which constructs a dataset based on the measured data asthe input.The BiGRU and the CNN utilize the strong correlation between neighboring rows of the image toextensively collect redundant information from the past and future inputs,to extract the stripe noise and to integrate thespatial information,and to utilize the difference between the data for loop iteration of this process.The model isvalidated through a large number of experiments and the BiGRU-CNN method outperforms other tested(classical)methods by reducing the vertical gradient energy by 15.2%and the residual nonuniformity by 21.9%.