The Research Of Adaptive Region Proposal Network And Self-Attention Mechanism Based Natural Scene Text Detection

Texts in natural scenes have rich semantic information and play an important role in automatic driving,robot navigation,automatic translation and other fields.How to extract text information in natural scenes accurately and efficiently has become one of the hottest issues in computer vision.This task includes two sub-tasks: text detection and text recognition.The former is the prerequisite of the latter.This paper studies text detection sub-tasks in natural scenes.The main work is as follows:(1)This paper will adopt an algorithm based on Mask R-CNN framework model,because it can adapt to the change of the shape and rotation of text objects in natural scenes.However,the original Mask R-CNN method has more post-processing steps,which will reduce the efficiency,but also affect the performance.To solve this problem,this paper adopts a text detection method based on corner regression,and only uses quadrilateral nonmaximum suppression to post-process the detection results.Compared with original Mask R-CNN method,this method has fewer post-processing steps and higher text detection performance.(2)In the original Region Proposals Network(RPN)of Mask R-CNN,a series of Anchors with different aspect ratios need to be set artificially during training.Because of the large dynamic range of the aspect ratio of text targets in natural scenes,it is difficult to cover these ranges with pre-defined sizes,resulting in a small number of high-quality positive samples for training RPN,which leads to a low recall rate of text targets in the testing stage.To overcome these shortcomings,this paper proposes a text detection method based on adaptive region recommendation network,which effectively alleviates the above problems and ultimately improves the recall rate of text targets.(3)The original Mask RCNN only relies on the region of interest to judge the category of the target.The lack of context information will lead to the algorithm often misjudges those text-like background as the text target.To solve this problem,this paper introduces the idea of self-attention mechanism into text detection task.This method can effectively construct the context information of candidate regions,thus effectively suppressing False Positives,and improving the accuracy of text target detection.The performances of all the proposed methods mentioned above have been verified by the experiments on benchmark datasets such as ICDAR 2015,MSRA-TD500 and so on.