Research On Feature Optimization And Pattern Recognition Method Based On Deep Convolutional Neural Network

Pattern recognition is an important branch of artificial intelligence.It allows the machine to observe the environment and learn how to judge different patterns.The quality of the feature is the key to the performance of the pattern recognition system,so a general feature optimization method is needed.This study is based on the two aspect of the optimization problem of pattern recognition,on the basis of using Deep Convolutional Neural Network(DCNN)to analyse,and study how to do feature optimize on different practical problem,thus to improve the recognition rate,and understand the network's behavior.The main research contents include:(1)In the analysis of two classification feature optimization method based on DCNN,combine with specific examples of two classification motor imagery in the brain computer interface,put forward a method named the Common Spatial Pattern(CSP)adaptive feature optimization based on DCNN.Based on the preprocessing of the original EEG signal,the corresponding feature matrix is obtained by the CSP transformation.Apply DCNN on feature matrix,and analyse the convergence of the DCNN network's full connection layer weights,according to the characteristics of network learning and get a CSP matrix feature selection criterion,obtained EEG efficient feature set F,computing the scale of F to construct CNN classifier.We worked on the BCI2005 IV a competition data set to perform experimental tests,get a recognition accuracy of 88.3%.The method compared with the improved method of CSP method,s CSP and KLCSP are tested on the same data set,and the average recognition accuracy is increased by 3.2%and 2.4% respectively.(2)On the basis of the two classification problem,the feature optimization method on multi-classification problem based on DCNN is analyzed.Aiming at the problem of constructing effective system in speech signal recognition,a feature reduction method based on DCNN is proposed.This method extracts the feature of the original speech emotion data,applies the DCNN to study the feature matrix,extracts the weights and defines the feature selection criteria MCFR-DCNN according to the network learning characteristic,calculates the difference of the activation weights of each class,and obtains the dimension emotional cognitive feature set F.In the multi-modal emotional database CHEAVD provided by the Institute of Automation,Chinese Academy of Sciences,we extracted all eight kinds of emotional data and tested the method.The use of all feature sets to construct the DCNN classifier reduced the error rate by 2.1%compared with the baseline,the feature set F obtained by this method is 9.4% lessthan the baseline error rate.This method only uses the 15% feature of the original feature,which not only reduces the complexity of constructing the actual speech emotion recognition system,but also reduces the recognition error rate.(3)In the further analysis of research,a complex feature optimization method based on DCNN weight matrix is proposed based on the analysis of complex feature multi-classification problem.After using DCNN to study the original data,the weight of the whole connection layer is extracted,and the feature value matrix is defined according to the network learning characteristic,and the original data and the characteristic value matrix are linearly combined to obtain a new data set.Experiments were performed on the large-scale image database Image Net,using three classifiers,which improved the recognition rate using the feature-optimized data set relative to all data sets.This paper studies different characteristics for different forms of data,the characteristics of learning DCNN network optimization and characteristics of the two dimension,provides a new method for feature identification and optimization problems in the field of pattern recognition.