EMANet: An Ancient Text Detection Method Based on Enhanced-EfficientNet and Multidimensional Scale Fusion
Top Cited Papers
- 5 July 2024
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11 (19) , 32105-32116
Abstract
Digitization through the Internet of Things is a key measure for preserving ancient books. As a crucial aspect of digitization, text detection plays an essential role in the preservation and dissemination of ancient culture. Nevertheless, common detectors for ancient books struggle to meet the dual demands of speed and accuracy required by the Internet of Things. Moreover, the robustness of these text detectors to capture complex features still needs to be strengthened. To tackle these issues, we introduce a new detector called EMANet. Firstly, we substituted the feature extraction module in the core of EfficientNet-B3 with an improved Feature extraction (MBConv++) module to better capture the dependencies between channels. By incorporating this module, the network can concentrate on crucial features within the entire datasets. This significantly contributes to fulfilling the requirements for both speed and accuracy in ancient text detection. Additionally, we devised a Multi-Dimensional Scale Fusion (MDSF) module, effectively bolstering the scale robustness of the network. Finally, we construct a mobile ancient text digitization app suitable for the Internet of Things. The proposed EMANet achieves an f-measure of 94.9% on SHAD datasets, demonstrating its effectiveness. Simultaneously, we evaluate the generalization capability of EMANet using the public MTHv2 datasets. Results reveal that EMANet outperforms the majority of existing text detectors. Furthermore, our model demonstrates exceptional computational speed and precision in detection when deployed within the Internet of Things framework, offering significant contributions to the holistic digitization of ancient texts.Keywords
Funding Information
- National Natural Science Foundation of China (61901059, 51978079)
- China University Industry–Academia–Research Innovation Fund (2023YC0061)
This publication has 42 references indexed in Scilit:
- Classifier-guided multi-style tile image generation methodJournal of King Saud University Computer and Information Sciences, 2024
- Domain adaptive multigranularity proposal network for text detection under extreme traffic scenesComputer Vision and Image Understanding, 2023
- Precise detection of Chinese characters in historical documents with deep reinforcement learningPattern Recognition, 2020
- Weakly supervised precise segmentation for historical document imagesNeurocomputing, 2019
- TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes2018
- SSD: Single Shot MultiBox Detector2016
- Detecting Text in Natural Image with Connectionist Text Proposal Network2016
- A knowledge-based recognition system for historical Mongolian documentsInternational Journal on Document Analysis and Recognition (IJDAR), 2016
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- Reading Text in the Wild with Convolutional Neural NetworksInternational Journal of Computer Vision, 2015