Research On Image Classification Of CNN-based Broad Learning System

With the rapid development of computer Internet technology,more and more people research and explore big data and artificial intelligence,so that these contents become the focus of the times.Among them,deep learning networks in the field of artificial intelligence can solve many complex pattern recognition problems by simulating the human brain,and tap the potential value of massive data.Even though deep learning networks are very powerful,there are still shortcomings.It is mainly reflected in: First,most of the existing deep network structures are complex,involving a large number of hyper parameters,and are plagued by extremely timeconsuming training processes;Second,the depth model adjusts parameters or continuously increases the number of network layers in order to obtain higher accuracy.Increasing the network layer is accompanied by the generation of a large number of parameters.The broad learning system born from this problem is simple in structure,fast,flexible and efficient,and can be an alternative method for deep learning networks,but there is a problem that the accuracy of image classification is not high enough.In addition,electronic equipment are widely promoted and popularized.It is of great value to timely extract,process and apply effective information generated by these products.Therefore,this thesis proposes and implements solutions to these problems as follows:(1)A CNN-Based broad learning system is proposed.First,the image passes the convolutional layer and the pooling layer to obtain the extracted features.In order to extract the features more accurately,the Adam algorithm is used to update the parameters of the convolutional layer and the pooling layer a number of times,Keep the output of the last iteration of the pooling layer and use it as input to the broad learning system.Then,the extracted image features are input into a broad learning system,and feature nodes and enhancement nodes are established.Finally,using the ridge regression algorithm to find the inverse matrix can solve the output weights.According to different data sets,different numbers of convolutional layers and pooling layers can be selected.(2)A lightweight network structure of CNN-Based broad learning system is proposed.On the one hand,the Squeeze Net design principle that reduces the number of convolution kernel parameters and the number of input channels,and removes the fully connected layer to simplify the parameters.On the other hand,the SVD algorithm is used to perform singular value decomposition on the feature nodes and enhanced nodes to reduce the dimensionality.In the field of image classification,compared to using a convolutional neural network or a broad learning system alone,the model constructed has both advantages,which improves the rate and accuracy,and is a new network that achieves a balance in terms of time and accuracy.Using the Squeeze Net design principles and the SVD algorithm to compress the overall proposed network,a lightweight network is finally obtained.