基于孪生胶囊网络的图像相似度度量技术研究

Research on Image Similarity Measurement Technology based on Twin Capsule Network

介绍了孪生胶囊网络,这是一种新颖的体系结构,将胶囊网络扩展到成对学习,从而实现图像相似度度量技术。模型以两个样本为输入,输出其嵌入高维度空间的表征,比较两个样本的相似程度。设计思路结合了胶囊网络,以记录信息更丰富的胶囊神经元代替传统神经网络的标量神经元,通过胶囊特征间的对比损失来训练模型。实验表明,孪生胶囊网络在小数据样本训练任务中的表现出色,能够完成相似度度量任务。

The Siamese capsule network is introduced in this paper,which is a novel architecture.The capsule network is extended to pairwise learning to realize image similarity measurement technology.The model takes two samples as inputs and outputs the representation of their embedding in high-dimensional space,and compares the similarity of the two samples.The design idea combines the capsule network,replaces the scalar neuron of the traditional neural network with the capsule neuron with richer recording information,and trains the model through the contrast loss between capsule features.Experiments show that twin capsule network performs well in the training task of small data samples and can complete the task of similarity measurement.