Deep capsule network for recognition and separation of fully overlapping handwritten digits

• Published in: Computers & electrical engineering Vol. 91; p. 107028
• Main Authors: Yao, Hongge, Tan, Yuxing, Xu, Chunqiu, Yu, Jun, Bai, Xiaojun
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 01.05.2021; Elsevier BV
• Subjects: Classification; Convolution; Deep capsule network; Digits; Dual routing cooperation; Feature extraction; Handwriting recognition; Overlapping digit recognition; Overlapping digit separation; Overlapping handwriting recognition; Overlapping handwriting seperation; Parameters; Routing iteration optimization; Separation; Overlapping handwriting seperation; Dual routing cooperation; Routing iteration optimization; Overlapping handwriting recognition; Deep capsule network; Overlapping digit recognition; Overlapping digit separation
• ISSN: 0045-7906; 1879-0755
• DOI: 10.1016/j.compeleceng.2021.107028

•Recognizing fully overlapped handwritten digits.•The dynamic routing algorithm between capsules is improved.•Feature extraction using small convolution kernel.•It has higher performance and lower parameters than capsnet. The recognition and separation of fully overlapping handwritten digits is an effective means to test the recognition ability of the network. It is also the basis of separating overlapping complex handwritten characters. This paper constructs a deep capsule network, FOD_DCNet, for the recognition and separation of fully overlapping handwritten digits. Firstly, we used small sized convolution kernels to extract features, which is conducive to extracting fine-grained features while reducing training parameter; secondly, we expanded the capsules dimension to express the extracted features to avoid the loss and omission of features; thirdly, we proposed "series dual dynamic routing collocation" to optimize the routing classification function. Compared with CapsNet, our FOD_DCNet reduces the number of iterations of each route, and increases the classification efficiency. The classification accuracy of FOD_DCNet can reach 93.53%, which is 5.43% higher than CapsNet and its parameter amount is only 55.61% of CapsNet. [Display omitted]