Research And Implementation Of Text Detection And Recogniion Algorith For Billboard Scenebasedon Gradient Segmentation

OCR(Optical Character Recognition)is a technology that recognizes the text information in the image and displays it on the computer.It is widely used in daily life.At present,the main research directions in the field of OCR are scene text detection and scene text recognition.The scene text has various forms,among which the billboard text has arbitrary shape(irregular,multi-directional,etc.),which brings great challenges to detection and recognition.For arbitrarily shaped text,the current mainstream research method in the field of scene text detection is based on segmentation,but the existing methods are difficult to distinguish adjacent text effectively,and the detection effect is poor,while the main research method in the field of scene text recognition is based on sequence recognition+attention,but the existing methods have the problem of attention drift and character angle skew,resulting in the recognition effect is not ideal.In view of the above problems in the field of scene text detection and scene text recognition,the main research contents of this paper are as follows:(1)In view of the problem that adjacent text is difficult to distinguish in the field of scene text detection,this paper takes the central area of the text area as the key factor of adjacent text differentiation.First,a gradient/weighted segmentation map(visually in the form of a heat map)is designed to accurately locate the central area of the text area.After the central area is predicted,it starts from the central area and expands to the complete text area according to certain post-processing steps,So as to get a complete text instance.(2)Aiming at the problem of attention drift and character angle tilt in the field of scene text recognition,this paper decomposes the recognition problem of two-dimensional level into one-dimensional level,first designs a sequence recognition+attention model on one-dimensional level,and then parallels two sequence recognition+attention models on x-axis and yaxis.It can accurately locate the character position on the one-dimensional level,so as to deal with the problem of attention drift.At the same time,the combination of the characteristics of the x-axis and the y-axis to predict characters can effectively deal with the problem of character angle skew and produce the effect of angle correction.