改进胶囊网络在图像识别中的应用

Application of Improved Capsule Network in Image Recognition

图像识别是指利用计算机对图像进行处理、分析和理解,以识别各种不同模式的目标和对象的技术,并对质量不佳的图像进行一系列的增强与重建技术手段,从而有效改善图像质量。本文用改进胶囊网络对MNIST数据集进行训练。胶囊是一组神经元,其活动向量表示一种特定类型的实体的实例化参数,它的长度代表实体存在的概率,方向代表实体的实例化参数,低层的活性胶囊,依据转移矩阵对高层胶囊的实例化参数进行预测,当多个预测一致时,高层胶囊被激活。本文利用spread损失来代替margin损失,避免过早出现“失活”胶囊,并且在不添加重构子网络的情况下,对不同路由迭代次数进行研究,确定路由迭代次数对分类准确率的影响,并确定模型最优参数。研究表明该模型在未做增强和扩展处理的MNIST数据集上的误分率低至0.32%。同时,改进胶囊网络在Fashion-MNIST,CIFAR-10数据集上也表现出了良好的性能。

Image recognition refers to the technology of using computer to process, analyze and understand images in order to identify targets and objects in different modes. And carry out a series of enhancement and reconstruction technical means for the poor quality image, so as to effectively improve the image quality. In this paper, the improved capsule network is used to train MNIST data set. Capsule is a group of neurons, and its activity vector represents the instantiation parameters of a specific type of entity. Its length represents the probability of entity existence, and its direction represents the instantiation parameters of entity. For low-level active capsule, the instantiation parameters of high-level capsule are predicted according to the transfer matrix;when multiple predictions are consistent, the high-level capsule is activated. In this paper, spread loss is used to replace margin loss to avoid premature “inactivation” capsule, without adding reconstruction sub network, different routing iteration times are studied to determine the impact of routing iteration times on classification accuracy and determine the optimal parameters of the model. The research shows that the misclassification rate of the model on MNIST data set without enhancement and expansion is as low as 0.32%. At the same time, the improved capsule network also shows good performance on Fashion-MNIST and CIFAR-10 data sets.