Research On Intelligent Accumulator Based On Convolutional Neural Network

Optical Character Recognition is a technique for scanning text on an electronic device to translate the content on the text into computer text.With the advent of the era of big data,the data generated by human beings is growing exponentially.How to extract the information that users are interested in effectively and quickly from texts,images and videos has attracted more and more researchers' attention.The traditional optical character recognition technology is based on artificial hand-designed features,and has poor character recognition for complex backgrounds,illumination,and distortion.In recent years,with the emergence of a new generation of OCR technology with deep learning technology as the core,the optical character recognition effect has made a huge breakthrough.Deep learning technology can learn the most robust classification information from massive data,which greatly improves the accuracy of image recognition and target detection.However,this technique relies heavily on large amounts of data to correct convolutional neural network parameters and improve model generalization capabilities.In this paper,a set of automatic accumulation algorithm is designed by using deep convolutional neural network for the handwritten digit detection and recognition tasks in the answer sheet.The algorithm attempts to use the small data set to complete the recognition task,and at the same time improve the generalization ability and detection speed of the model by optimizing the model hyperparameters and structure.A correction algorithm is designed to detect the digital tilt problem,which improves the reliability of the detection algorithm.The main work of this paper is as follows:(1)Aiming at the problem that the traditional multi-digit handwritten digit recognition algorithm can't divide the continuous number and the image distortion caused by the preprocessing process,the recognition accuracy is low.CRNN algorithm is used to train a model that can recognize multi-digit handwritten digits end-to-end.No pre-processing is needed,and it is proved by experiments that the recognition accuracy of the algorithm is much higher than that of the traditional algorithm,and it can effectively identify the continuous number.(2)Using the annotation tool,a printed text handwritten digital test data set containing 500 images was created and converted into a VOC dataset format.For the detection task,a one-stage detection algorithm SSD and a two-stage detection algorithm Faster-rcnn are respectively trained.By comparing the difference between the speed and precision of the two algorithms,the SSD algorithm is selected to detect the handwritten digits of the answer sheet area.(3)The K-means algorithm is used to analyze the target size and proportional distribution in the dataset,optimize the parameters of the SSD algorithm,and improve the generalization ability of the model.And through the experiment,the model structure is simplified and the detection speed is accelerated.Aiming at the problem that the detection picture may have the inclination to affect the recognition accuracy,the least square method is used to complete the automatic correction of the oblique picture,which improves the robustness of the algorithm.(4)Using the trained CRNN recognition model and SSD detection model,a paper automatic answering algorithm for the answer sheet is designed,and the automatic accumulation algorithm is debugged on the PC.An APP that can call this algorithm on the mobile side is designed to make the algorithm use more diverse scenarios.