基于改进主成分分析网络的手写数字识别方法

Method of handwritten digit recognition based on improved PCANet

针对人工特征对于多样性的变化没有很好鲁棒性,并且主成分分析网络提取的特征维数过高导致分类效率低且对内存消耗大的问题,提出了一种主成分分析网络和压缩感知结合的手写数字识别方法.首先,利用改进的主成分分析网络对图像进行特征提取;然后,用一个非常稀疏且符合压缩感知RIP条件的随机测量矩阵对抽取的特征空间投影,得到一个低维压缩子空间,该子空间可以保留高维图像特征空间的信息;最后,采用支持向量机对降维后特征进行训练和识别.实验结果表明:该方法识别率高、训练时间短;同时,该方法在加入池化层并进行压缩感知后,特征维数更低,模型内存占用更小,分类识别的速度更快.

The artificial features are not very robust to the variety changes,and the high dimension of features extracted by the principal component analysis network(PCANet) leads to low classification efficiency and high memory consumption.For solving these problems,a handwritten digit recognition method combining PCANet and compressive sensing was proposed.In this method,firstly,the feature of the image was extracted by the improved PCANet,and then a very sparse random measurement matrix which satisfied the restricted isometry property condition of compressive sensing was used to project the extracted feature space to obtain a low-dimensional compressed subspace.The information of high-dimensional image feature space can be well preserved in the low-dimensional subspace.Finally,the support vector machine was used to train and recognize the features whose dimension has been reduced.Experiments show that this method has high recognition rate and short training time.At the same time,after adding the pooling layer and compressive sensing,the feature dimension is lower,the model’s size is small,and the speed of recognition is fast.