Design And Implementation Of Document Image Recognition System Based On Convolutional Neural Network

With the continuous advancement of the smart city management system,the detection of urban manhole covers is a key link in the smart city management system.The lack and damage of manhole covers will cause serious traffic accidents and unpredictable economic losses.It can be seen that the detection and identification of urban manhole covers is very important.The detection of manhole covers must have accurate manhole cover information.When accurate manhole cover information has been obtained,these would precise information,The information is fed back to the corresponding department for the maintenance.If it is entered through manual inspection,for example the problems low efficiency,time-consuming,and low accuracy will occur.The problems is so difficult that we would not to complex natural scenes and high similarity between roads and manhole covers through traditional image algorithms.This paper improves the current mainstream image instance segmentation model algorithm,explores the instance segmentation model algorithm suitable for complex natural scenes,designs an urban manhole cover detection system based on convolutional neural network,and realizes the detection and recognition of smart city manhole cover images.A deep learning instance segmentation model and image multi-scale segmentation algorithm that integrates attention mechanism is proposed to detect urban manhole covers.After the information of the manhole cover is detected in the image,the specific information of the manhole cover can be obtained through OCR text recognition.These specific information are output in the form of text.The main tasks of the thesis are as follows:(1)Designed and implemented a small sample image enhancement method for manhole cover images,and proposed a data set based on urban manhole covers,which includes a web crawler for manhole cover images and a small sample image enhancement method.Manhole cover images are preprocessed and filtered out as part of the training data set.(2)Manually label the manhole cover data set obtained by Labelme labeling software,and split the data set into three parts: training set,test set,and verification set for subsequent image instance segmentation model learning and training.(3)Design and improve the current mainstream image instance segmentation model algorithm,solve the inapplicability of traditional image algorithms to images with rich natural scenes,realize accurate and rapid detection of manhole cover images,and lay the foundation for subsequent extraction of text on manhole cover images,Significantly improve the accuracy of manhole cover image character recognition.(4)Designed and investigated the text recognition method of the CTPN algorithm,and verified the accuracy of the CTPN algorithm text recognition method through the OCR open text platform and improved the recognition effect,which improved the accuracy of text detection and recognition to a certain extent.(5)Developed a set of system platform suitable for urban manhole cover detection,combined with some small samples of manhole cover data set enhancement in this article,manual Labelme annotation verification,construction of image instance segmentation model with channel attention mechanism module and CTPN,The algorithm combined with the character detection and verification of the OCR text opening platform has realized the batch and automated process of urban manhole cover detection.In addition,the system platform also includes the function of manual automatic verification,which can further improve the recognition accuracy and precision rate of city detection.