An algorithm for multi-directional text detection in natural scenes

Abstract

Due to factors such as background interference and scale variations, the text detection task in natural scenes is faced with challenges, especially in applications like autonomous driving and image understanding, where higher requirements are imposed on detection accuracy and efficiency. Under this background, the development of efficient detection algorithms specifically for natural-scene text is of particular importance. To this end, the Multi-Directional Text You Only Look Once (MDT-YOLO) is proposed in this paper. Firstly, a Dual Path Residual Connection (DPRC) block is designed, which enhances the model's multi-scale feature perception ability and alleviates the problem of missed detections caused by scale variations. Secondly, to reduce text information loss during downsampling, the Depthwise Separable Strided Downsampling (DSSDown) module is proposed, improving the model's ability to recognize fine - grained text regions. Additionally, an Efficient Down-Transition (EDT) module is constructed to reconstruct the Backbone network, achieving a coordinated improvement in semantic modeling and computational efficiency. Experimental results show that compared with the baseline model, the parameter count of MDT-YOLO is reduced by 29.4% while the processing speed remains basically the same. Meanwhile, on the MSRA-TD500 dataset, Precision and [email protected] are improved by 3.2% and 2.8% respectively, and on the HUST-TR400 dataset, they are improved by 1.7% and 1.3% respectively. The code will be available at https://github.com/WDP-0806/MDT-YOLO.