Enhanced Chemical Structure Recognition and Prediction Using Bayesian Fusion

In this paper we present a novel system for chemical symbol recognition capable of recognizing and predicting handwritten chemical structures. While the system is aimed primarily at chemists and search engine developers, it is also intended to be used as a learning tool in chemistry education on a student's touch screen computer or in classroom using smart boards. As a student draws on a device screen, the system would provide instantaneous feedback concerning possible chemical structures to complete the drawing, providing a continuously updated table of closely matching candidate symbols, and automatically identifying parts of the molecule that may be combined with other chemical structures. Major building blocks of the suggested system are two Bayesian-based image recognizers: one for collecting input images and one for producing the final output. Between input and output sketch classification is done using state of the art methods. By jointly incorporating feature-based and image-based techniques the system is able to recognize messy artifact-infected sketches and to predict possible terminations. Various performance related experiments proved the system to be considerably faster and more accurate especially when compared to neural-network based techniques.