Research On Arbitrary-shaped Scene Text Detection Algorithm Based On Deep Learning

In the field of computer vision,natural scene text detection is an important research topic.Since texts in natural scenes contain rich semantic information,natural scene text detection plays a fundamental role in many practical applications and has become a research hotspot in recent years.Text detection methods based on object detection and semantic segmentation will face the problem of contour modeling of the text object because of the complexity of the background and arbitrariness of text,such as distinguishing different adjacent text instances and accurately classifying pixels in text edge regions.Compared to the representation of text prediction results,the model’s ability to extract features from text edge regions has more impact on the accuracy of the contour model.Aiming at the problems faced by natural scene text detection,this paper has carried out the following work based on the semantic segmentation framework:(1)A text detection method based on edge feature enhancement is proposed.Because text detection methods based on semantic segmentation are prone to misclassification in the edge regions of text,edge regions have more impact on the accuracy of natural scene text detection.To improve the accuracy of text detection,this paper explicitly adds a text edge region detection branch to the model and pays more attention to features related to text edge in the feature extraction process through two feature enhancement modules.(2)A network module that enhances the foreground features is proposed to improve the feature extraction ability of the text detection model.In order to improve the representation ability of features by reducing the interaction between the foreground features and the background features,the proposed method divides the image features into features from the foreground region and the background region and then enhances them separately.At the same time,to further improve the accuracy of text detection,a text recognition module is added to the detection model.Compared with the basic model,experiments on scene text datasets have proved that the text detection accuracy,recall rate,and F-Measure of the two improved scene text detection methods proposed in this paper have been significantly improved.