Research on scene text detection algorithm based on modified YOLOv5

Abstract

Aiming at the problems of low precision, slow speed and the detection problems when the text lines are arranged in any direction in the traditional natural scene text detection method, based on the target detection algorithm YOLOv5, a rotated text detection method with angle classification is proposed——YOLOv5-R, by defining the representation of the rotating rectangle, calculating the method of rotating the IoU, and designing a new loss function to achieve accurate detection of horizontal and oblique text, and tested the effectiveness on the scene text datasets ICDAR2013 and ICDAR2015, after the transformation The algorithm realizes the function of rotating target detection, but there is still some room for improvement in arbitrary shape detection.