Digital Character Recognition Based On Machine Vision In Complex Background

With the era of big data,rich information is transmitted via scene images,and it is urgently needed to accurately extract useful information from this wealth of information in areas such as industrial automation,disabled services,transportation,and driverlessness.The textual information is particularly important as an important indicator language in the image.Due to the influence of light,background and other factors,the traditional identification method can not adapt to the complex and changeable environment.The positioning and identification of textual information under complex backgrounds have become the current research difficulties and focuses.With the application of neural network in the field of machine vision in the deep learning method,optical character recognition(OCR)as an important research direction in this field has also achieved rapid development.compared with the traditional method of artificially extracting image features,the neural network can automatically extract high-level features.Therefore,research on text recognition using convolutional neural networks has gradually become the mainstream of current research method.In this context,this paper focuses on the construction of a digital character recognition method based on convolutional neural networks in a complex context.Before digital character recognition,in order to make the acquired recognition image highlight the difference between the foreground and the background,several commonly used pre-processing algorithms are applied to preprocess images in complex background.Contrast analysis of the activation function,regularization method,and optimization method of the convolutional neural network.and comparisons are made between MNIST data and different optimization methods.It provides a strong basis for the selection of activation functions,regularization methods,and optimization methods.This paper divides the digit recognition in complex background into two tasks: character positioning and character recognition.The similarity between object detection and numeric character recognition is analyzed.The whole digital character is detected as a special "object".select the current part in the object The VGG-16 model in Faster-RCNN method with good characterization capability is used for feature extraction to generate convolutional feature maps;Complete the design of Region Proposal Net(RPN)and complete the target area selection;candidate is completed by Fast-RCNN network.Regional classification work,here is a two-classification process,the foreground class is a numeric character,the background class is a non-numeric character,and its loss function composition is analyzed;the recognition part is good for LeNet-5 which performs well on the MNIST data,The network is improved to complete the identification of foreground classes in the positioning network.Finally,Google's SVHN data set is selected as the basic data set.According to the specific requirements of the positioning and identification network,a corresponding data set is constructed to complete the training of the two parts of the network.The trained model was verified on the SVHN data set,and good results were obtained.