Algorithm Research Of Convolutional Neural Network Based On Domain Decomposition

The rapid development of Convolutional Neural Networks?CNN?has made major breakthroughs in the field of computer vision,and has made remarkable achievements in image recognition,image segmentation,and target detection.Convolutional neural networks have rich representation functions and can handle many complex problems in the field of computer vision.Problems such as automatic driving,target detection,and medical image imaging.As the scale of data increases,the number of layers of convolutional neural networks that cause processing problems deepens.Deep convolutional neural network training large-scale data sets takes a lot of time and takes up a lot of GPU computing resources.In practice,there are often not enough computing resources to deal with large data sets.How to deal with big data on small machines is very important in practice.This paper takes convolutional neural networks as the research object,and proposes a domain decomposition and combined transfer learning algorithm to solve this problem.This article mainly focuses on the two aspects of decomposition and combination of convolutional networks.This algorithm is inspired by the idea of domain decomposition,and proposes a decomposition and combination of transfer learning algorithms.Firstly,how to decompose and combine the convolutional network is introduced in detail.The second is to study how to divide the convolutional network,that is,how many subnetworks the convolutional network is divided into.Finally,the effectiveness of decomposition and combinatorial transfer learning algorithms is verified through experiments.The experimental verification part uses two common tasks of image classification and image segmentation in computer vision for verification,which reflects the wide applicability of the algorithm.For image classification tasks,this paper uses public data sets such as CIFAR₁0/CIFAR₁00 and Food-101.The image segmentation data is the original images of cardiovascular CT provided by each cooperative hospital.The experimental results show that the advantages of the convolutional network algorithm based on area decomposition have three aspects:1)The use of transfer learning strategy can accelerate the convergence rate of the network and improve the accuracy of image classification;2)The algorithm can handle large Scale data,and occupies few computing resources;3)The algorithm is also widely applicable to various convolutional network models and computer vision tasks.The main innovations of this paper are:?1?Split the deep convolutional neural network into multiple sub-networks,and use the weight combination of the sub-network training as the initial weight of the DCNN,which can provide a better initial training point for the DCNN In order to speed up the training of DCNN;?2?Each sub-network can train the corresponding sub-sample data independently and in parallel.DCNN pays more attention to the learning of local features through the sub-network initialization and improves the classification result of DCNN.