Research On Dot Characters Detection And Recognition Based On Image Analysis And Deep Learning

Image analysis and deep learning have made many significant achievements in the field of character recognition.However,there is still a lot of research space and value for special and specific character data set samples.This paper mainly studies the industrial dot character data set.Compared with other character data sets,the industrial dot character data set has the following characteristics:First dot character has less information than the continuous character,and the array of dot characters is more vulnerable to noise interference.Second reality of industrial dot character data set,each kind of data distribution is extremely uneven.This paper has done some exploration and research,the main work is as follows.(1)For the character data set of this article,the convolutional neural network optimizer,convolution kernel,and activation function are adjusted to improve the traditional character recognition Le-Net neural network,and construct a convolutional neural network suitable for the research object of this article.The improved neural network has increased the accuracy of character recognition in this article from 93.27%to 99.88%,and the convergence speed has increased by 20 epochs.(2)Aiming at the problem that dot matrix characters have less information and are susceptible to noise interference.This paper proposes a neural network character recognition method based on dilation.The specific method is as follows,first dilate the dot matrix characters to supplement the amount of dot matrix character information,then perform neural network training and recognition,Analyze and compare the character recognition accuracy and time efficiency of the sample after processing the same neural network structure and different dilation coefficients.(3)Aiming at the uneven distribution of each type of sample in the character data set,this paper proposes a neural network character recognition method based on character encoding information.The specific method is to divide the original 40-category network into 24 types of uppercase letters,6 types of lowercase letters,and 10 types of numbers through character encoding information to perform character recognition,and compared with traditional methods and data augmentation methods in terms of accuracy,time efficiency,and convenience.Combining expansion and character encoding information methods,the accuracy of character recognition increases from 99.88%to 99.95%when the recognition time is not much different.In the 16,000 test set,the number of recognition errors drops by 11.