Research On Handwritten Chemical Equation Recognition Based On Improved CRNN Model

The application of Chemical equation is everywhere in our daily life,affecting all aspects of us.As the basis of chemical industry,Chemical equation plays an important role.In recent years,with the continuous deepening of chemical research,a large number of related documents have emerged.Because Chemical equation is a two-dimensional structure that can only be saved as a picture,it takes up a lot of space.After its recognition and one-dimensional coding,it not only saves space,but also facilitates storage and transmission.In addition,deriving new Chemical equation is an important part of chemical research.Scientific researchers summarize the chemical components before and after the reaction through experiments,and obtain Chemical equation through derivation and other methods.An important step in the derivation of Chemical equation is balancing,while balancing through computers is relatively troublesome.If you balance on paper,you can use pens of different colors to mark at any position,which makes it easier to get a complete Chemical equation,and then input it into the computer efficiently through photo identification.In the field of education,it can be applied to the grading of chemistry exams and assignments.After scanning the exam papers and recognizing them as text sequences,computers can efficiently grade them.One of the most important goals in enterprises is to increase profits,and reducing product costs is one of the effective ways to reduce profits.In recent years,the price of memory modules has remained high.Reducing the space occupied by network model parameters can be carried on Edge device with smaller memory,thus reducing production costs.Therefore,the research goal of this article is to lightweight the model identifying Chemical equation,which is convenient to carry on lightweight equipment.This paper mainly does the following work for handwritten chemical equation recognition:(1)To solve the problem that there is no ready-made Chemical equation data set on the Internet at present,this paper has constructed a handwritten Chemical equation data set,collected 170 common Chemical equation and ion equations,asked different volunteers to copy each Chemical equation ten times,obtained 1700 pictures,and then expanded the data set with Gaussian blur of radius 1 to obtain 3400 images,Divide the dataset into training and testing sets in a 9:1 ratio.In addition,in order to solve the problem that the chemical way cannot be directly recognized due to the two-dimensional code,a new Chemical equation coding way is proposed,which converts the two-dimensional code into the expression of one-dimensional code for the convenience of later identification and marking.In terms of coding,we first classify different elements in the Chemical equation,including chemical molecular formula,reaction symbol,superscript,subscript,etc.,and code different coding methods.These coding methods are one-to-one,which can convert the coded structure into two-dimensional structure and facilitate its display.(2)In response to the problem of network parameters being too large to fit on lightweight devices,this paper proposes a new lightweight network model LCRNN based on the text recognition model CRNN(Convolutional Recurrent Neural Network),which uses the Mobile Net V3 M model in its convolutional layer.In addition,the recurrent layer of the model uses Bi GRU to further reduce model parameters and accelerate training speed.Through experiments,it was found that when using Bi GRU in the loop layer,the APc(character accuracy)of VGG,Mobile Net V3,and Mobile Net V3 M in the convolutional layer were 99.65%,90.00%,and 97.51%,respectively.When using Bi LSTM in the loop layer,the APc(character accuracy)of VGG,Mobile Net V3,and Mobile Net V3 M in the convolutional layer are99.61%,92.88%,and 97.69%,respectively.The experimental results indicate that Mobile Net V3 M can achieve similar results as VGG with fewer parameters.(3)In order to further improve the recognition accuracy of the LCRNN model,this article adds multi-scale fusion to the LCRNN model and names it MSF-LCRNN.This model uses two different downsampling methods to obtain two different scales,and fuses the larger scale upsampling with the smaller scale to complement each other’s advantages and improve recognition accuracy.On the Chemical equation data set,the APc(character accuracy)of LCRNN is 97.51%,while the APc(character accuracy)of MSF LCRNN is 98.34%.