基于胶囊网络的动态路由研究与改进

胶囊网络具有弥补卷积神经网络空间信息丢失和旋转不变性差的优点，已被广泛应用于图像分类、目标检测以及文本检测等多个领域，但胶囊网络仍存在参数量大且分类精确度低的问题。提出基于点乘注意力图卷积路由的胶囊网络分类模型。在同级胶囊之间构建连通图，通过注意力机制获取胶囊间的依赖关系，利用影响因素大的预测胶囊进行特征聚类，改变使用迭代更新高低胶囊层间耦合系数的动态路由方式，降低参数量并提升模型的分类准确率。此外，在特征提取部分加入残差网络提取更高维的特征以优化胶囊质量，在提升模型特征表达能力的同时可抑制模型过大。实验结果表明，在参数量小于多个胶囊网络变体的情况下，该模型在MNIST、FashionMNIST、CIFAR10和SVHN数据集上的精度分别达到99.74%、95.02%、91.78%和95.65%，均高于MS-CapsNet、TextCaps、AR CapsNet、FSc-CapsNet、DA-CapsNet等对比模型。

Research and Improvement of Dynamic Routing Based on Capsule Network

Capsule Network(CapsNet) has the advantage of compensating for the loss of spatial information and rotation invariance in convolutional neural networks.Therefore, they have been widely used in many fields such as image classification, target detection, and text detection.However, capsule networks still face some problems, such as having a large amount of parameters and low classification accuracy.This paper proposes a novel capsule network classification model that uses the dot-product attention mechanism to construct the graph convolutional layer, thereby selecting high-quality features to facilitate correct classification of the image.To obtain the spatial dependence between capsules within the same layer, the dot-multiplication attention mechanism is used to construct a fully connected graph between capsules of the same levels.The predicted capsules with large influencing factors are then used for feature clustering.In this way, the proposed model can reduce the number of parameters and improve model performance.At the same time, the residual network is added to the feature extraction process for the extraction of higher-dimensional features in order to optimize the quality of the capsule, which can not only improve the feature expression ability of the model but also restrain its size.The experimental results prove that when the number of parametersis less thanthat the number of multiple capsule network variants, the proposed model achieves 99.74%, 95.02%, 91.78% and 95.65% accuracy on the MNIST, FashionMNIST, CIFAR10, and SVHN datasets, respectively, making the proposed model more accurate than MS-CapsNet, TextCaps, AR CapsNet, FSc-CapsNet, and DA-CapsNet models.