Detecting Natural Scene Text Based On Convolutional Neural Network

With the development of deep learning,scene text detection has been greatly developed.However,many detection algorithms are still facing great challenges due to the different shapes,sizes,complex background and many kinds of languages in the natural scene text.Most of the early deep learning detection algorithms are based on bounding box prediction,and they do not pay attention to the regional change characteristics of text,which makes it difficult to separate closely spaced text instances.In addition,it is also difficult to detect various shapes of text,such as quadrilateral text and curve text,etc.To solve these problems,this paper explores the text detection algorithm of natural scene without preset anchor.The main research work is listed as follows.This paper proposes a scene text detection algorithm based on corner detection.The algorithm can detect text by locating four corners of the bounding box directly,and it does not need to preset any anchor.With the help of hourglass network and corner pooling prediction module,four detection branches output the heat map which is used to locate corners,the displacement which is used to scale the correction resolution and the embedded vector which is used to combine corners.On ICDAR 2015 and MSRA-TD500 datasets,the scene text detection algorithm based on corner detection achieves 80.7% and79.0% F-measure respectively.The experiment results show that it can significantly improve the performance of detection for horizontal text and quadrilateral text.In this work,the corner detection algorithm is difficult to detect curve text,thus we propose a 2D progressive mask kernel to describe the progressive variability of text region.It transforms the original label of the bounding box GT into the probability distribution map GT with a 0-1 progressive change.At the same time,a new progressive region prediction network(PRPN)with direction pooling module is proposed to predict the probability distribution of text region.The post-processing algorithm is used to convert the probability distribution of text region into the bounding box results,which can be used to detect scene text.Compared with the existing methods,this method has higher robustness and accuracy,without the design of bounding box,and is simple and effective to restructure.The experiments show that the method has achieved or is superior to the state-of-the-art methods in precision and recall.The method obtained 86.0% and 81.4%F-measure on ICDAR 2015 dataset and SCUT-CTW1500 dataset respectively.