LightDASS: a lightweight deformable transformer for unified single-point spotting on intelligent edge devices

Unstable illumination conditions such as strong reflections in maritime environments and environmental interference in meter reading scenarios severely degrade optical text recognition performance and intensify the trade-off between accuracy and computational efficiency. Existing methods struggle to simultaneously achieve robust recognition and lightweight deployment under such challenging conditions. To address this issue, we propose the lightweight deformable attention guided single-point spotter (LightDASS), an end-to-end optical character spotting framework designed for intelligent edge devices. The proposed method is built upon three key contributions. First, we reformulate text spotting as a sequence prediction problem, unifying detection and recognition as discrete token generation via an auto-regressive transformer. Second, we introduce a single-point annotation paradigm for training, which replaces conventional bounding-box supervision and significantly reduces labeling cost while maintaining competitive learning capability. Third, we design a lightweight transformer block with multi-head deformable attention and a group-wise feed-forward network, enabling adaptive focus on informative regions with substantially reduced computational overhead. We evaluate LightDASS on two challenging scenarios, including ship license plate recognition (SLPR) under severe maritime reflections and automatic meter reading (AMR) under low-light conditions. To further validate the generalization capability of the proposed method, we additionally conduct experiments on the ICDAR2015 dataset. Experimental results across all benchmarks demonstrate that our method consistently achieves superior performance over existing baselines, particularly in small-scale and degraded text scenarios while maintaining high efficiency. Furthermore, the proposed framework is lightweight and well-suited for real-time deployment on edge devices, enabling practical integration into intelligent systems for low-latency decision-making.