光电卷积神经网络的光学实时训练算法

Optical In-situ Training Algorithm for OptronicConvolutional Neural Network

摘要:

为了减少电子计算机中数字硬件的运算量,提升光电卷积神经网络(optronicconvolutional neural network,OPCNN)在实际应用中的分类性能,提出了一种新型的光学反向传播算法,在光学平台上实现了OPCNN 的实时训练,用于合成孔径雷达的目标分类。根据所提算法对OPCNN 进行训练,在训练过程中,网络正向传播与反向传播的运算操作均可以通过光学计算的方式以近光速的处理速度实现。基于运动与静止目标获取与识别(movingand stationary target acquisition and recognition,MSTAR)数据集的目标分类实验验证了所提光学实时训练算法的可行性。

Abstract:

In order to reduce the computational cost of electronic hardware and improvethe actual classification performance of optronic convolutional neural network (OPCNN), anoptical back propagation algorithm was proposed, and the in-situ training OPCNN in opticalplatform for synthetic aperture radar (SAR) target classification was realized. According tothe proposed algorithm, the main computation operations in the forward and backwardpropagation processes were carried out in optics, which with the speed of light. Experimentson moving and stationary target acquisition and recognition (MSTAR) datasets demonstratethe feasibility of the proposed optical in-situ training algorithm.