Text Detection And Recognition For Curved Targets

The style of the surface text on the curved target is different from horizontal text,multidirectional text and curved text,and it has certain geometric characteristics.In the detection stage,the bounding box label of the surface text on the curved target needs more vertex coordinates to express its spatial information,which brings difficulties to label a large number of training samples.In the recognition stage,it is more difficult to recognize the surface text on a curved target than regular text on a plane.Aiming at the task of surface text detection and recognition for curved targets,a text detection algorithm which is Text Mask Network(TMN)and a text recognition algorithm which is CPP-CRNN based on cylinder-plane projection are proposed in this thesis.TMN is a text detection algorithm which belongs to weakly supervised learning algorithms.The loss function of TMN is mainly divided into two parts: one is the weighted regression loss,the other is the sparsity constraint loss.The direction of gradient optimization of the two parts is opposite.TMN is a weakly supervised learning algorithm,which is reflected in that it does not require bounding box labels or text pixel labels of text pictures,but only needs background pictures corresponding to text pictures as supervised information.It solves the problem of labeling a large number of training samples.Comparative experiments show that the text detection accuracy of TMN has increased by4.5%.CPP-CRNN is a text recognition algorithm which is based on projection transformation.The pipeline of CPP-CRNN can be divided into two stages: projection transformation and text recognition.In the projection transformation stage,it applies a learnable cylinder-plane projection transformation sub-network(CPP)to process the candidate text regions generated by the text detection algorithm.CPP transforms cylinder text into plane text.In the text recognition stage,firstly,a convolutional neural network extracts the basic features of the text image and then a bidirectional recurrent neural network for feature fusion extracts the context features of the character sequence,and finally the recognition result is produced.Comparative experiments show that the text recognition accuracy of CPP-CRNN has increased by 3%(character accuracy)and 8%(word accuracy),respectively.According to the characteristics of actual medical application scenarios,a medical tag intelligent identification system is preliminarily built with the adoption of the proposed text detection algorithm(TMN),text recognition algorithm(CPP-CRNN),the correction algorithm for recognition results and super-resolution reconstruction algorithm.The system mainly contains an input module,a detection module,an identification module,an auxiliary identification module and a display module.The proposed text detection algorithm(TMN)and text recognition algorithm(CPP-CRNN)are adopted for the detection module and the identification module,respectively.A super-resolution reconstruction algorithm is adopted in the auxiliary identification module to enlarge the text area three times to display which can assist the anesthetist in visual identification.The system is designed to assist anesthesiologists in pre-operative preparation and reduce the anesthesiologist's medication error rate.At present,the medical tag intelligent identification system can identificate 18 commonly used anesthetic drugs,but this system still needs to be improved before it can be applied in actual medical scenarios.