基于改进ALEXNET卷积神经网络的电容层析成像三维图像重建

Three-Dimensional Image Reconstruction of Electrical Capacitance Tomography Based on Improved ALEXNET Convolutional Neural Network

针对卷积神经网络三维图像重建算法的样本训练速度慢和成像精度低的问题,提出一种根据不同流型的AlexNet神经网络数据训练方法。首先通过SVM算法将输入的电容样本数据按照流型分类,然后采用单一流型样本数据训练相应的AlexNet卷积神经网络,使得某一流型的神经网络的输入样本数据类型简单、样本数量少和神经网络规模小。同时采用具有冲量和自适应学习速率的Adam算法,减少了训练时的误差振荡,加速神经网络的收敛。通过对比改进的AlexNet卷积神经网络算法和LBP算法的成像结果,表明优化后的AlexNet在成像精度和速度上有显著提升。

A method is proposed that the corresponding AlexNet neural network is trained according to the data of different flow patterns for the problem of slow sample training and low imaging accuracy for the three-dimensional image reconstruction algorithm of convolutional neural networks.The input data is classified by SVM according to the flow pattern,and the corresponding AlexNet convolution neural network is trained for single-class sample data,which make the input data type simple of the neural network,the number of samples and the neural network is small.The AlexNet convolutional neural network uses the Adam algorithm with impulse and adaptive learning rate which can reduce the error oscillation and accelerate the convergence of neural networks during training.By comparing the imaging results of the improved AlexNet convolutional neural network and the LBP algorithm,it shows that the optimized AlexNet has a significant improvement in imaging accuracy and speed.