Handwritten Experimental Digital Recognition Based On Convolutional Neural Network

With the continuous progress of science and technology,the informatization of various industries continues to improve,and digital has become an important carrier of information transmission,playing a huge role in industrial production and daily life.However,there are still a lot of work and experiments that cannot be simulated in the computer.A large amount of data exists in the form of handwriting,but a large amount of data often needs to be analyzed by the computer,and manual data collection and entry will consume a lot of manpower and material resources.Therefore,how to accurately recognize handwritten experimental numbers becomes particularly important.With the rapid development of deep learning,image recognition and image classification technologies based on deep learning are becoming more and more mature.Therefore,this thesis adopts a method based on convolutional neural network to perform handwritten experimental digital recognition.In this thesis,by combining traditional image processing,deep learning and machine learning related technologies,on the basis of research and learning of existing text detection algorithms,a handwritten experimental digital recognition algorithm based on convolutional neural network is proposed,including data positioning,Number segmentation and number recognition three core steps.In the detection and extraction of experimental data,this thesis builds a CRAFT network model,and produces corresponding data sets to train the network model,and at the same time designs a post-processing algorithm for the data in this thesis,so as to achieve effective detection and extraction of experimental data,Tested on the ICDAR 2017 dataset and obtained a detection accuracy of 79.8%.In the digital segmentation part,this thesis uses the water drop method to segment the numbers;in the digital recognition part,this thesis designs a high-precision handwritten number classification network for the problem that the handwriting recognition accuracy rate cannot meet high precision.Firstly,the continuous asymmetric convolution is used to extract the initial features of the image while reducing the parameters required for the calculation.Secondly,the depth separable convolution is used to improve the Inception structure,and the residual network is combined to prevent gradient dispersion.Finally,the Softmax classification is performed.Through the MNIST data set experiment,a recognition rate of 99.45% is obtained.In order to further improve the network recognition rate,the support vector machine(SVM)is used in the classification layer to replace the fully connected layer and the Softmax layer of the traditional convolutional neural network(CNN).After cross-validation,a recognition rate of 99.78% is obtained.Finally,the handwriting experiment numbers in this thesis are tested for recognition,and a recognition rate of 99.18% is obtained.The results show that the improved Inception structure can obtain a larger network width.At the same time,SVM also has a good effect on the classification ability of CNN extracted features.