Research On Feature Extraction Algorithm Based On Convolution Principal Component Analysis

With the development of artificial intelligence and machine learning,feature extraction and selection techniques are widely used in the fields of image processing,pattern recognition and data mining.In pattern recognition,feature extraction and selection is a very important step,feature extraction and selection directly affect the effect of classification recognition.Feature extraction and selection is a commonly used dimensionality reduction algorithm,the most effective feature to express the high dimension of the original data,can eliminate the redundancy between the data to achieve the feature space dimension compression.Principal component Analysis(PCA)is an important feature subspace learning method,which allows the original sample point to be projected into a low dimension space through a linear transformation,thus reducing the number of data dimensions.In one image,the neighboring pixels are more closely connected,while the pixels relative to the distance are weaker.The convolution neural network utilizes the spatial characteristic of image pixel,the connection mode of the convolution layer and the previous layer is the local connection and share weight value,which greatly reduce the number of connection parameters in the network and enhances the learning efficiency of the network.At present,the common feature extraction methods are based on the global characteristics of the whole image,and do not make full use of the spatial characteristics between pixels.In view of the above problems,a new method of face feature extraction based on convo-lution is proposed,which is inspired by convolution neural network.The method is to block the image,use the PCA-L1 method to obtain the projection matrix of each block position,s-elect the convolution template to extract the local feature,and then combine the local feature to extract the global feature.The final extraction is based on local-global characteristics.This method makes full use of the characteristic of image space pixel,which has both rotational in-variance and translational invariance.Experimental results on AR and ORL database show that this method can extract the features better.