A shared-backbone detection-recognition framework for end-to-end industrial character reading

In industrial manufacturing and quality traceability, accurate extraction of character information from metal nameplates is essential for automated recognition. This paper proposes an end-to-end industrial character extraction framework with a shared backbone, termed Shared-Backbone Industrial Character Recognition Network (SBICR-Net). Within a unified feature space, the framework jointly performs character region localization and content recognition. In the detection branch, Faster Implementation of Cross-Stage Partial Bottleneck with Two Convolutions with Asymmetric Padding (C2f-AP) replaces the standard Faster Implementation of Cross-Stage Partial Bottleneck with Two Convolutions (C2f) in the backbone, thereby enhancing boundary perception and local texture modeling for small-scale and densely distributed characters. In the recognition branch, the proposed Group-Level Orientation Rectifier (GLOR) and the improved Strip-wise Large Kernel Attention (LKA-S) are introduced to unify sequence orientation and strengthen local structural representation, and are combined with Scene Text Recognition with a Single Visual Model (SVTR) for sequence recognition. Experiments on a real-world railway wheelset metal nameplate dataset show that, under the End-to-end setting, the proposed method improves character-level recognition accuracy and word-level recognition accuracy by 7.8 and 30.9 percentage points, respectively, while reducing the number of parameters by 23.9% and increasing inference speed by 16.5% compared with the baseline. Under the Recognition-only and Detection-only settings, the proposed method achieves performance improvements of varying degrees over advanced models. In addition, supplementary experiments on the public Metal Parts’ Surface Character Dataset (MPSC) and Chinese City Parking Dataset 2020 (CCPD2020) further demonstrate its applicability across different character scenarios.