Transformer-Augmented Pointer Detection Network (TAPDN) For Accurate Analog Gauge Reading In Industrial Environments

Abstract

Analog gauge meters remain critical in industrial monitoring but are challenging to read accurately in noisy, low-light, or occluded environments. Manual inspection is inefficient and error-prone. This study proposes an automated system capable of accurate and real-time analog pointer meter reading in complex conditions. We propose the Transformer-Augmented Pointer Detection Network (TAPDN), an advanced deep learning architecture that synergizes EfficientNetV2 and Swin Transformer backbones for robust local-global feature extraction. TAPDN incorporates a Multi-Scale Attention Fusion (MSAF) module, a dual-head decoder to simultaneously localize pointer tips and estimate orientation, and an adaptive multi-task loss function for effective joint optimization. Evaluated on the Pointer-10K dataset, TAPDN achieves state-of-the-art results with 95.0% OKS AP and 76.3% VDS AP, outperforming baseline models while running at 32 FPS. TAPDN offers a robust and scalable solution for intelligent industrial inspection, effectively handling low-quality inputs and supporting real-time deployment in smart manufacturing environments.