Research On Post-processing Of Scene Text Detection

Text is a carrier of information and plays an important role in life.In recent years,with the development of artificial intelligence,people increasingly hope that machines can automatically recognize the semantic information of text in text images,and to recognize the text in text images,we must precisely locate the text in images,that is,complete text detection to get the location information of text in text images,so text detection gradually becomes a research hotspot in the field of computer vision.Generally speaking,text detection methods are mainly divided into traditional feature-based text detection methods and deep learning-based text detection methods.The traditional feature-based methods use manually designed features to verify the obtained candidate regions and finally obtain the prediction boxes coordinates.With the rise of artificial neural networks and deep learning techniques,deep learning-based text detection methods tend to achieve better detection results.Therefore,deep learning-based text detection methods gradually become the mainstream text detection methods,while traditional feature-based text detection methods gradually fade out of people's view.Deep learning-based text detection methods are generally divided into two steps: obtaining candidate boxes by detection models and post-processing the obtained candidate boxes.The post-processing usually adopts non-maximal value suppression or prediction box merging connection and so on.Current text detection methods based on deep learning can often achieve good detection results.However,due to the diversity of text scenes and the quality of the image itself,many current text detection methods still have problems such as character misrecognition and detection deviation.Character misrecognition is to detect some useless non-text information mistakenly as text,so as to obtain redundant prediction boxes,such as punctuation marks,some patterns in complex scenes,etc.The detection deviation is to get a large gap between the coordinates of the final prediction box and the real label,and a single text is easy to get multiple prediction boxes after detection,so that the text cannot be completely framed by the text box.Therefore,in order to solve the above two problems,this paper starts from the perspective of post-processing,and proposes post-processing methods of text detection for character misrecognition and detection deviation respectively.Specifically,the main contributions of this paper include the following two aspects:(1)To address the problem of character misrecognition problem often encountered in the text detection field,this paper proposes a non-standard box removal algorithm from the perspective of post-processing of text detection to improve the accuracy of text detection.In general,we think that in the same text image,the size of the characters and the characters should be roughly the same.The design of the non-standard box removal algorithm is mainly based on this principle.The algorithm is to determine whether each prediction box belongs to the standard box by the given prediction box area threshold.If the condition is satisfied,the prediction box area is smaller than the maximum area threshold and larger than the minimum area threshold,the prediction box is judged to belong to the standard box,and if not,the prediction box will be eliminated.What remains at the end is the updated set of prediction box coordinates,that is,the processed detection result.This method can improve the accuracy of text detection on the basis of a slight drop in text detection speed.The effectiveness of this method in improving the performance of text detection is verified through the comparative analysis of experimental results on three datasets: canonical Yi,Chinese2 k,and English2 k.(2)To address the problem of detection deviation in text detection,this paper proposes a text detection post-processing method for detection deviation to improve the performance of text detection.The method is mainly divided into three modules.The first is the background removal module.The background information and text information are separated by pixel threshold.Since different text detection scenes require different pixel thresholds,in order to adapt to different detection scenes,we propose Adaptive pixel threshold algorithm,this algorithm can calculate the text pixel threshold of the prediction box from the image information extracted by the prediction box.Then there is the candidate box expansion module,which expands the prediction box in all directions by determining whether there are text pixels at the boundary of the prediction box.Finally,the coincidence box removal module is used to remove the completely coincident and incompletely coincident prediction boxes in the final detection result obtained by the algorithm.At this time,the detection result obtained is the detection result processed by the method in this paper.This method can improve the accuracy,recall and F1 value of text detection while still maintaining a high detection speed.Finally,we have verified the robustness and time complexity of the method in this paper on six data sets including standard Yi language,Chinese2 k,English2k,ICDAR 2013,ICDAR 2015 and ICDAR2017(CTW-12k)through experiments.And compare them with LSAE,EAST,CTPN,TextBoxes,SegLink,TextBoxes++,CRAFT and other text detection models are compared,and the results show that the method in this paper can improve the performance of text detection.In summary,based on the current text detection methods,this paper proposes two post-processing methods,which solve the problems of character misrecognition and detection deviation frequently encountered in text detection,and make a certain contribution to the field of text detection.At the same time,a large number of experiments have proved that the method in this paper can effectively improve the performance of text detection.