Research On Text Detection And Recognition Based On Deep Learning

Text detection and recognition have attracted much attention as hot research topics in the field of computer vision.They have been widely used in work and life,such as document identification,license plate recognition,industrial production automation.They complement each other to form a complete optical character recognition(OCR)system.Text detection is the premise and basis of text recognition,which is used to find the position of the text line from a large image,Text recognition is used to recognize the extracted text line,and outputs the final result of the system.They determine the final performance of an OCR system.In the past,text detection and text recognition were basically implemented based on artificially designed features and traditional image processing methods.These features and algorithms are difficult to be designed and require a lot of prior knowledge,which will result low accuracy and low generalization.In recent years,deep learning has been developed rapidly,and successfully applied in the field of computer vision such as image classification,object detection,and semantic segmentation.Deep learning-based algorithms,as data-driven algorithms,automatically discover and learn the characteristic rules implied in a large amount of data through iterative training and without excessive human intervention.So,deep learning has better performance compared with traditional image processing algorithms.In this thesis,detonator coded character detection and recognition are studied based on deep learning,aiming to solve a practical industrial problem efficiently and accurately.The main work of this thesis includes:(1)For the collection and processing of detonator data,this thesis designs a detonator data acquisition system and a series of automatic data annotation tools.The data acquisition system collects detonator image data efficiently and safely,and the automatic annotation tool can greatly improve the work efficiency of training data preparation.(2)For text detection,this thesis proposes a detonator coded character detection network,termed as MFCNNet,based on multi-fully convolutional network fusion.The proposed network incorporates attention mechanisms,the fusion of multiple network information,and an improved loss function.Attention mechanism effectively suppresses the interference of background noise;the fusion of multiple network information makes multiple sub-networks complement and promotes each other;the improved loss function effectively suppresses unnecessary noise responses around the text area.Experiment results indicate that these improvements have effectively improved the accuracy of the proposed network.The detection precision and recognition accuracy reach 99.835% and 98.026%,respectively,representing the most advanced level.(3)For text recognition,this thesis proposes a detonator coded character recognition network,termed as FICNN based on flip-invariant convolution kernel.FICNN is a network that uses flip-invariant convolution kernels on the basis of the general recognition network.The flip-invariant convolution kernel effectively improves the processing capability of the network for flipped images.This is a solution for the phenomenon of detonator image flipping.Experiments show that FICNN improves the accuracy of the detonator text recognition by 5.24% and 1.09% compared with the classical methods,which achieves the best performance and proves its effectiveness.