深度EM胶囊网络全重叠手写数字识别与分离EI北大核心CSCD

基于胶囊网络的向量神经元思想和期望最大算法(Expectation-maximization,EM),设计了一种以EM为向量聚类算法的深度胶囊网络(Deep capsule network,DCN),实现了重叠手写数字的识别与分离.该网络由两部分组成,第1部分是“识别网络”,将EM算法改为EM向量聚类算法,以替换原胶囊网络CapsNet中的迭代路由部分,这一改动优化了网络的运算过程,实现了重叠数字识别.第2部分是“重构网络”,由结构完全相同的两个并行网络组成,对双向量进行并行重构,实现了重叠数字的分离.实验结果显示,对于100%全重叠手写数字图片本网络识别率达到了96%,对比CapsNet在80%的重叠率下95%的识别率,本文网络在难度提升的情况下,识别率有明显提高,能够将完全重叠的两张手写数字进行图片进行准确地分离.

Fully Overlapped Handwritten Number Recognition and Separation Based on Deep EM Capsule Network

Based on the idea of vector neuron of capsule network and expectation maximization algorithm (EM), a deep capsule network (DCN) with EM as the vector clustering algorithm is designed to recognize and separate overlapping handwritten digits The network consists of two parts. The first part is “identification network”. The EM algorithm is changed to EM vector clustering algorithm to replace the iterative routing part of the original capsule network CapsNet. This change optimizes the network operation process and realizes overlapping number recognition. The second part is the “reconstruction network”, which is composed of two parallel networks with identical structure. The bi-vector are reconstructed in parallel to realize the separation of overlapping digits. The experimental results show that for 100% full overlap handwritten digit, the recognition rate of the network reaches 96%. Compared with the 95% recognition rate of CapsNet at 80% overlap rate, the recognition rate of the network in this paper is significantly improved in the case of increased difficulty, and can accurately separate two completely overlapping handwritten digits.