Research On Convolutional Neural Network Based On Two-Dimensional Wavelet Transform

Deep Learning plays an important role in improving the level of artificial intelligence.Among which,Convolutional Neural Network(CNN)is widely used in image classification,object detection and other fields because of its "Sparse Connection" and "Parameter Sharing" characteristics,which has good self-learning ability and robustness.As a data-driven and endto-end feature extraction technology,the model structure of the CNN has becoming larger and more complex with the rapid increase of all kinds of data samples.Moreover,this form of learning features from data also has a characteristic which is "black boxes".In other words,it is difficult for people to understand the logic of the prediction inside the model,which makes the model lack of certain interpretability.Therefore,the visual analysis of the CNN,the optimization of the model structure,and the combined research with the interpretability module,all help humans to understand its internal logic.these methods are also useful for convergence speed,model complexity,and recognition accuracy of the CNN model.Based on the discussion of the combination of the interpretability module and the traditional CNN,this thesis proposes a method of replacing the first-layer convolution kernel of the traditional CNN with a convolutional kernel based on two-dimensional wavelet transform.And in the scene of image classification,the influence of this method on CNN is verified.The main research work is as follows:(1)Based on the wavelet transform,this thesis performs discrete sampling of twodimensional continuous wavelets function,replaces the first convolutional layer of the traditional CNN with the wavelet convolutional module(WCM)designed in this thesis.And the first convolutional layer of the traditional CNN is replaced by WCM.The WCM is used to extract features of image.At the same time,the parameters of dilations and translations in WCM are manually set by the introduced selective initialization method.So,the Two-Dimension Wavelet Convolutional Neural Network(2D-WCNN)model is proposed.the data-set of Dogvs-Cat and PASCAL-VOC are used to verify the performance of the 2D-WCNN model in scene of image classification.This thesis specifically compares the influence of different mother wavelet functions and the size of the convolution kernel in WCM.The experimental results show that the 2D-WCNN model has better performance than traditional CNN.The accuracy has increased by up to 4.9% on the data set of Dog-vs-Cat.The accuracy has increased by up to 7.05% on the data set of PASCAL-VOC.(2)In order to make up for the shortcomings in the WCM that the parameters of dilations and translations are manually set and cannot be learned,and to optimize the time required for training,an adaptive wavelet convolutional module(AWCM)is proposed based on the backpropagation algorithm of CNN.The AWCM causes the parameters of dilations and translations to be updated by backpropagation.On this basis,the AWCM replaces the first convolutional layer of the traditional CNN,and reduces the number of its convolution kernel.Two-Dimension Adaptive Wavelet Convolutional Neural Network(2D-AWCNN)model is proposed.The great performance of 2D-AWCNN model in extracting image features is verified in the scene of image classification.The experimental results show that,the accuracy has increased by up to 5.76% on the data set of Dog-vs-Cat.The accuracy has increased by up to9.01% on the data set of PASCAL-VOC.The combination of two-dimensional wavelet transform and convolutional neural network based on the principle of interpretability module proposed in this thesis has verified in the image classification.The design of interpretability module provides clear semantics for CNN to a certain extent.Meanwhile,the experimental results have proved that it can effectively improve the performance of the model.The CNN model based on the two-dimensional wavelet transform proposed in this thesis has achieved great results on both the data-set of Dog-vs-Cat and PASCAL-VOC,with faster convergence speed and higher recognition accuracy.