面向手写数字图像的压缩感知快速分类

Compressive sensing fast classification for handwritten digital images

为了减少神经网络模型对手写数字数据集的训练计算耗时和最佳训练次数,同时保证手写数字图像的分类准确率,引入了压缩感知技术,提出了基于压缩感知和单隐层前馈网络(Compressive Sensing and Single Hidden Layer Feed⁃forward Network,CS-SHLNet)的手写数字图像快速分类算法。首先,利用高斯随机矩阵对具有稀疏性的手写数字图像进行线性观测,将高维图像信号投影到低维空间得到观测值;其次,通过误差反向传播(Error BackPropagation,BP)算法不断调整单隐层前馈网络权值建立适应于观测值的神经网络模型,将观测值嵌入神经网络中对图像进行特征提取;最后,采用单隐层前馈网络模型对手写数字进行图像分类,以训练计算耗时、最佳训练次数和分类准确率等指标对模型进行定量评估。实验结果表明:相比较单隐层神经网络和深度学习对MNIST手写数字数据集的高维图像信号图像分类,先通过CS技术利用观测数M=235的高斯随机矩阵线性观测得到图像的观测值,再利用单隐层前馈网络对观测值进行图像分类,网络模型的训练计算耗时缩短为13.05 s,最佳训练次数缩短为3次,分类准确率保持97.5%。该算法中的压缩感知线性观测可以有效减少神经网络模型对手写数字数据集的训练计算耗时和最佳训练次数,而且可以保证分类准确率。

To reduce the training calculation time and optimal training times of a neural network model and ensure high classification accuracy of handwritten digital images,compressive sensing technology was in⁃troduced,and a fast classification algorithm of handwritten digital images based on compressive sensing and a single-hidden layer feedforward network(Compressive sensing and single-hidden layer feedforward network,CS-SHLNet)was proposed.First,a Gaussian random matrix is used to obtain a linear measure⁃ment of the handwritten digital image with sparseness,and the high-dimensional image signal is projected to the low-dimensional space to obtain the measurement value.Second,using the error backpropagation(BP)algorithm,the weights of the neural network are continuously adjusted to establish a single-hidden layer feedforward network model suitable for the measurement values,which are embedded into the neural network for image feature extraction.Finally,a single-hidden layer feedforward network model is used to classify handwritten digits,and the model is quantitatively evaluated by the time-consuming training calcu⁃lations,the optimal training times,and the classification accuracy.Experimental results show that—in contrast to using a single-hidden layer neural network and deep learning for high-dimensional image signal classification of MNIST handwritten numeral datasets—through the CS technology,the Gaussian random matrix linear measurement number,i.e.,M=235,is first used to obtain the image measurement value;then,the single-hidden layer feedforward network is used for image classification.The training calculation time of the network model is reduced to 13.05 s,the best training times are reduced by a factor of three,and the classification accuracy is 97.5%.The compressive sensing linear measurement in the algorithm can effectively reduce the computation time of the training and the optimal training times of the neural net⁃work model for handwritten digital datasets and the classification accuracy can be ensured.