Rotation Invariance for Offline Handwritten Chemical Organic Ring Structure Symbols Recognition

Abstract

Auto grading is an instruction tool which could reduce teachers' burden, provide students with instant feedback and support highly personalized learning. One key technique is to recognize students' handwritten assignments. This work will be focused on the task of recognizing handwritten chemical organic ring structure symbols which exhibit the characteristic of rotational symmetry. Due to the global parameter sharing mechanism and pooling operation, convolutional neural networks (CNNs) have the power to learn translation-invariance features. However, the design of the standard CNN itself does not specifically consider the rotation invariance. In this paper, we explore different methods to improve the property of rotation invariance of CNN and apply them for semantic recognition of handwritten chemical organic ring structure symbols. The first one is a test-time augmentation method with voting strategy without modifications to the network architecture. For the second method, we add 2 new layers into the model to endow it with the property of rotation invariance. These 2 proposed methods are evaluated on a self-collected data set and achieve the recognition aecuracy of 98.75% and 93.125% respectively.