Research On Deep Learning Based Natural Scene Text Detection

Text in images provides rich and precise high-level semantic information,which is important for numerous potential applications such as scene understanding,image and video retrieval.As a necessary means of understanding scene text,text detection has attracted increasing attention in computer vision in recent years.However,due to the diversity of text scales and the uncertainty of image quality,text detection in natural scenes is still a challenging problem.Inspired by the successful application of convolutional neural network in object detection task,this paper proposes a text detection method based on multi-level feature extraction combined with deep learning technology.The proposed Multi-level Feature Extraction Module can capture more text details,thus enhancing the robustness of the algorithm to small-scale text and low-resolution images.In addition,the proposed method is end-to-end,and can directly output the text detection results.This paper evaluates the method on several standard data sets,and the experimental results prove the effectiveness of the method.To handle multi-scale texts in images,this paper proposes a novel scale-adaptive text detection method,which introduces a scale regression layer to learn the scales of texts in an image,and adjusts the sizes of priori boxes accordingly.Compared to other box-based text detection methods,the proposed method largely reduces the number of priori boxes and thus improves the computational efficiency.In addition,the scale-adaptive respective field is proposed to replace the standard respective field,which is widely used in the existing methods.The proposed algorithm can automatically adjust the sizes of the receptive fields according to the scales of texts in an image,so as to reduce the background interference and extract more necessary features in the feature extraction stage,and further improve the detection accuracy.In this paper,a comparative experiment is conducted with other state-of-the-art text detection algorithms on standard datasets.The experimental results show that the proposed method can significantly improve the computational efficiency of the box-based detection framework,and further enhance the robustness against multi-scale text,especially small text.