Text Detection Algorithm based on Improved YOLOv3

Focusing on scene text detection, this paper adopts an improved scene text detection algorithm based on YOLOv3 network. Firstly, for a single detection target, the training speed of Darknet53 backbone network used by YOLOv3 is slow due to too many layers, so that this paper proposes a method for replacing it by Darknet19. Secondly, multi-scale detection was retained in the original network, and three anchors of different sizes were used to predict the bounding boxes. The experiment results show that the training speed of this algorithm is greatly improved and the detection effect is accurate and stable.