Research And Application Of Multi-scale Capsule Neural Network

The capsule network is a novel neural network proposed in recent years that aims to solve the shortcomings of traditional convolutional neural networks,which cannot effectively learn the spatial relationships among entities in images.Currently,the capsule network performs well in the classification tasks of the MNIST and Fashion-MNIST datasets.However,since the research is still in its early phases,there are some issues that need to be resolved.Firstly,the poor performance of the capsule network on complex datasets suggests that there is still room for further optimization of the architecture and routing algorithm.Secondly,since the neurons of the capsule network extend by one dimension compared to the traditional neural network,it requires numerous matrix-matrix multiplication operations during the inference process.This places high demands on the deployment device’s memory bandwidth and matrix operation performance.As a result,further study is needed to determine the appropriate end-side deployment scheme for capsule networks.To address the issues mentioned above,this thesis refines the algorithms and theories of capsule networks in terms of network architecture design,routing algorithm optimization,deployment scheme design,and applications.The main innovative of this thesis is as follows:(1)To address the issues of poor performance and high computational cost of capsule networks with complex datasets,this thesis offers a software-hardware co-designed multi-scale capsule neural network model named MMRCaps Net.In terms of algorithms,MMRCaps Net constructs multi-level residual capsule modules to enhance the extraction capability of the capsule network;introduces atrous convolution to enhance the extraction capability of the contextual information;and proposes a reconstruction sub-network based on transposed convolution to achieve better results.Compared with the classical capsule network model Caps Net,MMRCaps Net reduces the number of parameters by about 50% through a lightweight architecture design,and improves the classification accuracy by 8.42% and 0.9% on the CIFAR10 dataset and SVHN dataset,respectively.Inspired by the idea of "neuromorphic computing",this thesis proposes a circuit implementation scheme for MMRCaps Net based on memristor crossbars from the application of capsule network deployment in end-side devices.The non-volatile properties of the memristor devices enable "storage and computation integration" with extremely low power consumption in this scheme,which supports parallel computation of matrix-matrix multiplication.This scheme is expected to advance the deployment and utilization of capsule networks in intelligent terminal devices.(2)To address the issues of inefficient routing algorithms for capsule networks,this thesis proposes the scaled dot product attention capsule network model named SDACaps Net,which is applicable to medical image classification tasks.In terms of network architecture,it eliminates the reconstruction subnetwork while maintaining excellent performance,and the overall number of network parameters is only 31% that of MMRCaps Net.The feature extraction module of SDACaps Net consists of several residual convolution modules to extract multi-scale feature information from images.In terms of routing algorithms,SDACaps Net introduces the attention mechanism and proposes the scaled dot product attention routing algorithm SDA Routing,which makes the transfer between feature vectors of each capsule layer more refined and accurate.Next,this thesis evaluates and analyzes the performance of SDACaps Net with four datasets containing two kinds of medical images.Several performance metrics show that SDACaps Net performs well in medical image classification tasks,especially on small sample datasets,which is of great practical value in responding to public health emergencies.Finally,this thesis compares the performance of SDACaps Net with other advanced capsule network models as well as several classical convolution neural networks to validate the effectiveness of SDACaps Net.This thesis investigates capsule networks in terms of network architecture,routing algorithms,deployment schemes,and applications.The findings contribute to the theoretical advancement of capsule networks and offer certain application values.