Chemical structure recognition and prediction: A machine learning technique

A chemical structure recognition system designed to identify and predict hand-drawn chemical diagrams is presented. Although the proposed system is primarily intended to help chemists complete their chemical structure drawings and predict and complete the chemical structure, it is also envisioned to be utilized as a learning aid in chemistry education on a student's touchscreen device or on a teacher's smart board. The proposed system continuously provides the user with real-time feedback about the most relevant combinations to complete the chemical compound. A constantly updated list of prioritized candidate structures is displayed to the user. If the user is an expert, he or she will be allowed to contribute to the learning process of the proposed system. The proposed system deploys two trainable Bayesian networks, one at its input and another one at its output. Between the two networks, the classification of hand- drawn chemical structure is done using morphological image processing, LDA and PCA, incorporating therefore both image- based and feature-based techniques. This incorporation contributes considerably to the increased performance of the proposed system by enabling it to efficiently recognize messy drawings and noisy sketches like those containing over-traced strokes. Performance comparison with existing state of the art techniques proved the system to be outperforming in terms of recognition accuracy.