Recognition Of Handwritten Chemical Organic Ring Structure Symbols Via Improving The Rotation Invariance Of CNN

Due to the development and progress of society,more and more chemical-related works have been transferred to smart devices.However,intelligent devices are still very unfriendly to the input of chemical formulas,which are inefficient and slow,bringing people an inconvenient user experience.In this case,the recognition of handwritten chemical formula is urgent.Therefore,the research topic of this thesis is to recognize handwritten chemical organic ring structure symbols.There are 36 classes of chemical organic ring structure symbols in this thesis,including 5 basic classes and 31 classes which are derived from 5 basic classes by rotation.In this thesis,36 classes of chemical organic ring structure symbols are recognized as 5 basic classes in semantic recognition.Therefore,the chemical organic ring structure symbol has rotation invariance.However,at present,the convolutional neural network only has translation in-variance and does not have the ability to recognize rotation invariance.Therefore,this thesis uses several methods to improve the rotation invariance of convolutional neural network,and applies them to improve the semantic recognition accuracy of handwritten chemical organic ring structure symbols.Firstly,the data augmentation technology is used,which is used on the training data and the test data respectively.Rotate each chemical organic ring structure symbol every 30 degrees,for a total of 12 times.After the test data was augmented,a voting mechanism was used to select the category with the most frequent occurrence in the recog-nition results as the final recognition category.Secondly,the structure of the convolutional neural network is improved.Two new network layers,the cyclic slicing layer and the cyclic pooling layer,are used and added to the input and output of the network respectively.The cyclic slicing layer and the cyclic pooling layer used in this thesis contain two structures of 4 rotation angles and 12 rotation angles.So that the convolutional neural network can learn the structure characteristics of different rotating copies of the chemical organic ring during training.The total number of images in the experimental data set in this thesis is 810,and all im-ages are divided into training set and test set according to the ratio of 4:1.The experimental results show that the data augmentation technique and rotational invariance enhancement technique used in this thesis have achieved good results in the semantic recognition experi-ment of handwritten chemical organic rings,and the two methods have achieved the highest recognition accuracy of 98.75% and 93.125% respectively.