RFANET-OCR: AN ATTENTION-BASED END-TO-END FRAMEWORK FOR DETECTION AND RECOGNITION OF TEXT ON METAL SURFACES

Industrial metal part images often suffer from low visual contrast, uneven illumination, surface corrosion, and clutteredbackgrounds—posing significant challenges for accurately detecting and recognizing marking characters. These conditionsfrequently impair existing methods, leading to poor localization of low-contrast text regions and degraded recognitionperformance. In this work, we build upon the Refined Feature-Attentive Network (RFANet) framework, which combinesregression-based and segmentation-based strategies for robust text localization. We retain its core architecture—includingparallel feature integration, attention map generation, and proposal refinement—and extend it by integrating an Optical CharacterRecognition module for end-to-end industrial text spotting. This addition enables both detection and recognition within a unifiedpipeline. We evaluate our enhanced method on two large-scale industrial scene text datasets—MPSC and SynthMPSC—originally introduced in the RFANet paper, comprising over 102,000 annotated images and 1.9 million text instances acrossdiverse backgrounds and character structures. Extensive experiments on these datasets and four public benchmarks demonstratethat our proposed extension achieves state-of-the-art performance in both text detection and recognition tasks.