Research Of Image Classification Algorithm Based On Convolutional Neural Network

With the advent of the cloud era,big data has attracted more and more attention.Classifying and sorting all kinds of information on the Internet,professionalizing and processing these meaningful data,and realizing the value-added of the data are the fundamental purpose.For the widely existing image information,we have no way to easily identify and classify its meaning by keyword search like textual information.Therefore,how to classify images by computer vision and pattern recognition is particularly important.Although the traditional pattern recognition method can be used to classify images,it also highlights some problems.In general,the accuracy of the recognition depends to a large extent on the features chosen by the person and it is empirical.In addition,these features are selected for specific identification objects,and they are difficult to apply to other categories of identification and classification.The convolutional neural network algorithm model can abstract the original image information layer by layer,gradually forming high-level features,and this process is completed by the network itself,and then repeatedly adjusts the weights to achieve the classification effect.This dissertation focuses on the application of convolutional neural networks in the field of image classification.Based on some of the latest achievements,the current convolutional neural network model is analyzed from many aspects and many improvements have been made.When training samples are insufficient or overtrained,overfitting is often caused.The intuitive manifestation is that as the training process progresses,the complexity of the model increases and the error rate on the training set gradually decreases,but the error rate on the verification set actually increases.Inspired by Dropout and DropConnect,this thesis proposes a new regularization algorithm that fuses the two.The concept of Boosting algorithm is to combine several weak classifiers into one strong classifier with higher performance through some specific methods.This thesis combines the convolutional neural network with AdaBoost algorithm to form the BoostCNN model.It can well combine multiple weak classifiers into one strong classifier to improve the performance of the algorithm.The convolutional neural network model has attracted much attention because of its ability to extract different levels of abstract features.The nonlinear activation function of each hidden layer plays a crucial role in the extraction of hierarchical features.In order to increase the feature representation capability of the activation function and overcome the problems of the traditional activation function in the deep neural network model,this thesis proposes a training multi-layer Maxout network activation function.Using this activation function not only improves the feature extraction andfeature representation capabilities of the neural network model,but also overcomes the optimization problems such as the disappearance of gradients occurring in the training process.In summary,this thesis compares the improved convolutional neural network algorithm with the traditional convolutional neural network algorithm.The results show that the improved algorithm can improve the accuracy of image classification.