Optimization Method Of Handwritten Digit Recognition Based On Deep Learning And Ghost Imaging

Handwritten digit recognition technology as an important branch of optical character recognition has a wide range of applications,such as bank statements,postal codes,census.These departments have relatively high requirements for handwritten digits,so the accuracy of its recognition also needs to reach a higher level,so how to efficiently recognize is particularly important.In recent years,the method to solve such problems mainly adopts the means of machine learning.Although this method can avoid the problem of low recognition efficiency caused by manual interpretation in some aspects,the method based on machine learning needs pretreatment,feature extraction and other processes.However,in the process of preprocessing,it is easy to cause the loss of image detail information,and in the feature extraction of handwritten digital image,the extraction of detail information such as edge and texture will be incomplete,which will affect the recognition efficiency.Thus,it is particularly important to jump out of the stereotype that identifying objects is identifying images.Ghost imaging,as a counter-intuitive indirect imaging method,does not need to obtain the details of the object in advance,and can realize the rapid imaging of the object without looking directly at the object.Therefore,combining it with automatic handwritten digit recognition technology can effectively solve the problem of low recognition efficiency caused by incomplete detail information.This paper organically combines computational ghost imaging with deep learning,and optimizes the performance of handwritten digit recognition system based on ghost imaging from two dimensions of classification recognition and image reconstruction.The main research contents of this paper are as follows:(1)Firstly,the relationship between font complexity and recognition effect of handwritten digits is analyzed.Then it is concluded that the key to improve the recognition problem of each kind of handwritten digits is to find a recognition method which can adapt to the complexity of the font structure.Then,by analyzing the problems existing in the traditional handwritten digit recognition and ghost imaging reconstruction algorithms,this paper leads to the optimization of handwritten digit recognition from the perspective of classification and reconstruction ideas.(2)An optimization method of handwritten digital classification and recognition based on ghost imaging: Firstly,a ghost imaging detection system was built on an optical experimental platform.In order to reduce the redundancy between speckles.Hadamard speckle field was designed to irradiate handwritten digital images and the total light intensity projected by handwritten digital was collected by using a bucket detector without spatial resolution;Then,based on the advantages of convolutional neural network in image classification,the network architecture was built and the values of the bucket detector were converted into one-dimensional vectors and the corresponding handwritten digital images were used as the input of the network for training;Finally,the experimental results show that the proposed method has higher recognition accuracy.In order to further verify the effectiveness of the proposed method,a comparative analysis is carried out with the fully connected neural network and the traditional CNN network.The method using the ghost imaging principle,only through the handwritten digital transmission of total intensity value can be quickly classified,boost the handwritten digital recognition efficiency.(3)In order to solve the problem that the information contained in the light intensity sequence at low sampling rate is not enough for the network to recognize handwritten digits with high accuracy due to the inherent limitations of the classification network,a generative adversative network model is improved from the perspective of reconstruction image optimization based on the most intuitive visual feeling.Firstly,a large number of light intensity sequences are obtained by second-order correlation between Hadamard matrix and target handwritten digit images,and corresponding original handwritten digit images are used as training sets to reduce the collection time of ghost imaging detection data and obtain sufficient training data.Secondly,it optimized and improved the existing generative adversarial network,and built a network model suitable for the field of ghost imaging reconstruction algorithm;Finally,experiments show that the proposed method can directly recover the target handwritten digital image from the one-dimensional light intensity sequence whose measurement times are much lower than the pixel value,so as to achieve the recognition of handwritten digital image under the condition of low sampling rate,and it is compared with the traditional ghost imaging reconstruction algorithm.It effectively promotes the practical process of ghost imaging technology and the application of deep learning in more fields.