A Real-Time Scene Uyghur Text Detection Network Based on Feature Complementation

Text detection in complex background images is a challenging task. With the national emphasis on the culture of various ethnic groups, carrying out the task of detecting Uyghur in natural scene images is of great significance to the intelligent information technology industry and economic construction. In fact, current text detection applications in complex scenes have large models, and sluggish detection rates, and are challenging to implement on mobile devices. Uyghur text has unique writing characteristics that lead to low detection results. To address these problems, we propose a feature-complementation-based text detection framework for real-time scenes (FC-Net). The network uses a deeply separable residual convolutional network (DResNet) for feature extraction, which lowers the model’s parameter count and boosts the detector’s detection speed. Secondly, a feature enhancement network based on spatial feature attention (SFA-FEN) is included in to obtain multi-target information by expanding the range of perceptual fields and enhancing the robustness of small-scale Uyghur. The experimental results show that FC-Net can maintain the advantages of a lightweight network while maintaining high accuracy, effectively reducing model parameters, and speeding up model detection. The robustness and real-time performance are achieved not only on the Uyghur dataset but also on an arbitrary text dataset.