Complex convolutional neural network: cv-cnn

Complex convolutional neural network (CV-CNN) Feature extraction of complex input data with complex parameters and variables , And feature classification . at present ,CV-CNN There are two main research directions ： One is about CV-CNN Theoretical research , The neural network in the real number field is now brilliant. A large number of new algorithms and models have been proposed , The theoretical innovation of algorithms based on the characteristics of complex numbers is a hot research topic . The second is CV-CNN Exploration in application field ,CV-CNN The emergence of has made the problems in many fields get new solutions , Explore new application scenarios extensively , And carry out adaptive innovation , It is also a research hotspot .
The basic structure of complex neural network mainly includes input layer 、 The hidden layer and the output layer . Layer to layer full connection , There is no connection in the layer . however CV-CNN The input data of is a complex number , The network weights and offsets are initialized in the complex form . In the process of forward propagation , The real part and the imaginary part of the weighted sum of input and weight are excited nonlinearly . therefore , The output of each layer of neurons is also complex . The backward training process of the network adopts the gradient descent algorithm in the complex field .
In the neural network algorithm of multilayer structure , The lower level and the higher level are respectively used to learn the low-dimensional and high-dimensional features of the target . The input layer is usually composed of width 、 Height and depth characterize , Depth indicates the number of channels of input data . stay SAR Image classification problem , The multi-channel complex image can be directly used as the input of the network .
The typical process of feature extraction in convolution network includes convolution 、 Nonlinear excitation and pooling . The convolution layer convolutes the input data with a learnable filter , Both input and output can be 2 D matrix . The convolution result generates the characteristic graph through the nonlinear excitation function . Nonlinear excitation functions are commonly used Sigmod、ReLU etc. . The next layer of the convolution layer is usually the pooling layer , Used for down sampling characteristic graph , Thus reducing network parameters , The properties of convolutional neural networks include local connections 、 Weight sharing 、 Pooling and cascading multiple layers . about CV-CNN, All structures of the network include convolution layer 、 The characteristic diagram of the pool layer and the parameters of the filter are complex .
notes ： The content of the article is extracted from 《 Synthetic aperture radar intelligent interpretation 》 Xu Feng et al