Abstract:With the advancement of industrial automation, detecting text on drum labels has become increasingly important for intelligent quality inspection. However, the cylindrical surface of drums introduces geometric distortions, along with non-uniform illumination and background clutter, which limit the performance of conventional scene text detection methods. To address these challenges, this study proposes a geometry-aware dual-branch normalized feature fusion framework for label text detection on industrial cylindrical containers. Specifically, a hierarchical pyramid backbone called P-VimNet is built upon Vision Mamba (Vim) to perform global modeling, and a geometry-aware module, PRSS (polar ring-scan sequencer), is designed in the high-resolution branch to align the principal direction and deformation path of the curved surface. During feature fusion, BiFPN weighted fusion is adopted and formulated as a fusion operator that takes two inputs and produces one output across adjacent scales. This operator replaces the fusion operations at each node in the network architecture diagram to improve the stability and adaptivity of cross-scale feature combination. Finally, a DBNet++ differentiable binarization head is employed to generate segmentation-based text regions. Experiments on a self-constructed drum label dataset in the tobacco blending field demonstrate that the proposed method achieves strong detection performance, with a precision of 91.3%, a recall of 90.1%, and an F1-score of 90.7%, while maintaining an inference speed of 44 f/s. The method also performs robustly on the public ICDAR2015 dataset, achieving a precision of 92.5%, a recall of 85.5%, an F1-score of 88.9%, and an inference speed of 43 f/s. These results indicate that the proposed approach delivers both robustness and real-time capability under complex industrial imaging conditions.