Research Of Statistical Modeling Based On Convolutional Neural Network

Deep learning has been applied more and more in the field of image processing.Its convolutional neural network model has shown high performance in tasks such as image recognition,classification and segmentation,which has attracted the attention of many scholars.While convolutional neural networks have good performance,they also have some shortcomings that cannot be ignored.On the one hand,there are a large number of weight parameters in the convolutional neural network model,which leads to the operation of the convolutional neural network model often has high requirements for computer hardware;on the other hand,the high accuracy of convolutional neural networks is often based on high-quality label datasets,and the acquisition of label information in datasets requires a lot of time and labor costs,and the labels of datasets are difficult to obtain in many practical problems.In order to solve these two problems,this thesis combines statistical methods and deep learning methods to explore new statistical deep learning methods,aiming to obtain image segmentation models with higher accuracy,fewer parameters and higher efficiency.This thesis mainly studies the following three aspects:（1）A spatial converter empty space pyramid（STASPP）network module is proposed.The module can adaptively extract the auxiliary information in the image and use the auxiliary information to effectively improve the segmentation accuracy of the small volume model.Based on the improved fully convolutional neural network（FCN）,the STASPP module is added to the FCN to obtain a high-precision STASPP-FCN model.Compared with the FCN model,the MIo U of the STASPP-FCN model increased from 93.98%to 99.23%on the simulated dataset and from 63.17%to79.99%on the DRIVE dataset.（2）A self-supervised learning method based on Dirichlet process mixture model is proposed,which can complete the training of the model without the help of artificial label information.On the basis of self-supervised training,the weight of convolutional neural network is sparse by using the idea of Lasso.This method compresses the weight of convolution kernel by adding L₂regularization term to the loss function to achieve the purpose of model sparseness.The model using Lasso weight sparsization method reduces the time required to segment one image on the CPU from 276ms to 270ms.（3）A model compression method based on factor analysis is proposed.This method uses factor analysis to reduce the number of convolution kernels and the number of channels.The weight tensor after dimension reduction is set as the weight of the compression model,so as to realize the compression of the model and ensure the accuracy of the compressed model.This method can compress up to 45.60%of the parameters on the FCN model and up to 36.71%of the parameters on the STASPP-FCN model.