Handwritten Chemical Equations Recognition Based on Lightweight Networks

Handwritten chemical equations recognition is one of the important research directions of optical character recognition (OCR) and text recognition technology, which is widely used in life. Although the mainstream deep learning text recognition model can get good recognition results, the number of parameters of the model is too large to be carried on some portable devices. We develop a new lightweight network model (LCRNN) based on the CRNN model for handwritten chemical equations recognition. Firstly, in the convolutional layer of the LCRNN model, we propose a new MobileNetV3 (MobileNetV3M) to reduce number of the model parameters. The MobileNetV3M changed the original down-sampling method to max-pooling, so it can extract more critical information. Secondly, we use the BiGRU model in the recurrent layer. Finally, a new chemical equations encoding method is proposed, which can change the two-dimensional chemical equation into one-dimensional chemical equation encoding, so as to facilitate handwritten chemical equations recognition. The experiments demonstrate that the character precision of the LCRNN model is 2.1% lower than the CRNN model, but the number of parameters is significantly reduced.