Capsule Network Framework for Handwritten Character Recognition

Abstract

A Capsule Network Framework for Handwritten Recognition Based on TensorFlow introduces the novel framework of Capsule Networks (CapsNets) from the TensorFlow system, in order to propose and implement a framework for reliable and accurate handwritten recognition in the architectural power of Capsule Networks. The processing of variable-layout, handwritten, traditional Convolutional Neural Networks (CNNs) is a great challenge because they fail to detect hierarchical spatial relationships and affine transformations. Capsules, which are collections of neurons that encode both the presence and pose (position, rotation, scale) of features, are used by CapsNets to solve this problem. Part-whole relationships are then generated with the help of a Dynamic Routing mechanism. The framework concentrates on important tasks when tested on the EMNIST Letters benchmark data set, which consists of 145,000 grayscale letters of the English alphabet (uppercase and lowercase). The TensorFlow implementation features a unique CapsNet layer and the training loop is designed to reduce the reconstruction loss and classification loss. Experimental results show that the proposed CapsNet framework has an improvement in generalization over the baseline CNN models with the increased recognition accuracy and robustness to handwriting variations within the 26 character classes. In this work, the authors have verified that CapsNets is a very successful pose-aware deep learning approach for document analysis and recognition in real world scenarios.