Open AccessText detection method based on HDBNet in natural scenes
Author(s) -
Wang Li,
Yao Xiang,
Song Chenwei
Publication year - 2023
Publication title -
the journal of engineering
Language(s) - English
Resource type - Journals
ISSN - 2051-3305
DOI - 10.1049/tje2.12212
Subject(s) - offset (computer science) , computer science , artificial intelligence , measure (data warehouse) , bounding overwatch , minimum bounding box , set (abstract data type) , differentiable function , pattern recognition (psychology) , binary number , text detection , precision and recall , image (mathematics) , mathematics , data mining , mathematical analysis , arithmetic , programming language
To solve the problem of large expansion offset of text detection in natural scenes, a text detection method based on HDBNet is proposed. First, the probability map of the text region is obtained by segmenting the image. The binary map is obtained by using the differentiable binarization of the probability map. The bounding box of the text region can be obtained by looking for the connected region on the binary map. Then, aiming at the problem of large expansion offset of text detection in natural scenes, a scheme of height prediction is adopted to compensate for the expansion loss caused by the width‐to‐height ratio. Finally, experiments show that when the depth network structure is ResNet‐50, the Precision, Recall and F‐measure of the proposed HDBNet method in TD500 data set are 0.9196, 0.8058 and 0.8590, respectively. Moreover, in MLT data set, the Precision, Recall and F‐measure of the proposed HDBNet method are 0.9393, 0.7884 and 0.8573 respectively. The results of HDBNet are higher than those of the comparison methods. Therefore, compared with the comparison methods, the proposed HDBNet method can significantly improve the performance of the text detection system.