Research On Scene Text Detection

Scene text detection has been widely used in many fields in real life,such as image retrieval,driver assistance,and industrial detection.The traditional methods mainly focus on the extraction of hand-crafted features based on the color,texture,and stroke width of the text.Such methods have weak generalization and detection performance.In recent years,deep learning-based text detection methods can better extract deeper features of images with a deep convolutional neural network,bringing better performance and generalization.After combing and summarizing the above methods,this paper improves on the shortcomings of existing CNN-based text detection methods,and designs three different text detectors.The specific works are as follows:(1)Research on scene text detection based on confidence fusion.As the last step of the text detection task,the non-maximum suppression(NMS)algorithm merges and filters the repeated detections of the same real text box.In this process,NMS uses the classification confidence of the prediction box as the basis for ranking,which may make appropriately those better-located but owing lower classification confidence text boxes suppressed,affecting the detection recall rate.This paper addresses the above deficiencies.Firstly,an intersection-over-union network is designed to predict the intersection-over-union(IOU)between each anchor box and the matched ground-truth and the IOU is regarded as the localization confidence.Secondly,in NMS,the classification confidence is replaced by the fusion of classification confidence and localization confidence as the ranking standard to preserve the accurately located text boxes.Thirdly,experiments on ICDAR2011 and ICDAR2013 show that the method can improve the accuracy of text detection and the detected text box is more compact and contains less background area.(2)Research on anchor-free horizontal scene text detection.Firstly,for the problems of anchor-based text detection in designing many hyperparameters for anchors,an unbalanced number of positive and negative samples,and an unrobust detection of narrow and thin text regions,an anchor-free text detection method is designed to directly predict the distances from each coordinate point on the feature map to the left,top,right and bottom of the real text box.Secondly,the dynamic feature selection strategy in model training is improved,and the gradient is passed only to the feature layer with the smallest loss function in each instance.Thirdly,experiments on ICDAR2011 and ICDAR2013 show that the method can effectively improve the speed of text detection and is more robust to small and long text regions.(3)Research on anchor-free arbitrary scene text detection.Firstly,the text in the real scene is usually skewed or distorted,so,in this paper,the text area is marked with a rotated rectangle with an angle.Secondly,for the problems of arbitrary text detection methods in many parameters related to angles and slow detection speed,an anchor-free arbitrary text detection method is proposed,which directly predicted the distances between each coordinate point on the feature map from the left,top,right,and bottom of the rotated rectangle,and angle from the long side of the rectangle to the horizontal right.Thirdly,experiments on ICDAR2013 and ICDAR2015 show that the method can shorten the detection time of each picture to 32% of the detection time of mainstream methods while ensuring the effectiveness of text detection in any direction.