Font Size: a A A

The Application Of Compression And Acceleration Of Convolutional Neural Networks Methods In Remote Sensing Image Classification

Posted on:2021-04-06Degree:MasterType:Thesis
Country:ChinaCandidate:F LiFull Text:PDF
GTID:2392330611963163Subject:Control engineering
Abstract/Summary:PDF Full Text Request
Deep CNNs have been applied to most fields and have achieved amazing results with the sharp growth of DeepLearning.Using CNN to extract and classify the depth features of remote sensing images can make the classification performance of remote sensing image scene classification tasks get a huge improved.Even though using deep cnns can improve the capability hugely.However,deep neural networks,especially deep convolutional neural network models(VGG16,ResNet,DenseNet,etc.)also have the problems of the amount of parameters,huge calculations,and storage overhead.Research on the lightweight method of convolutional neural network suitable for remote sensing image scene classification tasks is of great significance for the deployment of models in resource-constrained environments.According to this,the following work is mainly completed here:First,the background of the research was expounded,then we analyzed the domestic and international research status of remote sensing image scene classification tasks and neural network light-weighting.The relevant theoretical basis of convolutional neural networks is introduced in detail.At the same time,several lightweight network structures and two excellent coarse-grained pruning methods are mainly discussed based on the network structure design and coarse-grained pruning that this article focuses on.An improved framework for lightweight network MobileNetv2 is proposed.DenseNet's idea which is dense connection is applied to the lightweight network MobileNetv2,and feature reuse is used to improve network performance.Get a smaller model by using the combine of a bottleneck with an expansion coefficient of 1,a stride of 1 and a bottleneck with an expansion coefficient of 1,a stride of 2.At the same time adjust the number of bottleneck output channels.Get a smaller network model by using this method.The validity of the model is verified on the remote sensing image scene classification data set NWPU-RESISC45.A compound pruning algorithm for filter pruning of the network is proposed.Compound pruning is filter pruning and a coarse-grained pruning method from the pruning granularity.Compound pruning first defines the elasticity of the filter of the convolutional layer.The elasticity of the filter is used to measure the influence of the filter on the loss function to determine the importance of the filter.At the same time,the scaling factor of the BN layer after the convolutional layer is used to determine the importance of the feature map generated by the filter,and delete the filter when both are determined to be unimportant.In order to test the effectiveness of compound pruning,this paper performs compound pruning on a variety of commonly used network models on the benchmark dataset Cifar10 and compares it with other different pruning methods.In terms of remote sensing image scene classification tasks,this paper uses remote sensing image scene classification data set NWPU-RESISC45 to composite pruning two widely used networks,VGG16 and ResNet50,and compare the classification performance before and after pruning.Experiments show that composite pruning can basically maintain the classification performance while effectively reducing the amount of network parameters and calculations.At the same time,the comparative experiments before and after compound pruning also illustrate the effectiveness and practicability of the pruning method in this paper.
Keywords/Search Tags:Remote Sensing Image, Deep Learning, Model Compression, Neural Network Pruning, Convolutional Neural Network
PDF Full Text Request
Related items