Steel surface defect detection with a dual path alignment super-resolution network

In industrial production systems, images often suffer from low-resolution due to limitations in acquisition equipment, posing challenges for precise defect identification. Existing methods have rarely been explored for high-fidelity reconstruction on low-texture steel surfaces. To enhance image quality and defect detection performance, this paper proposed an improved dual-path fusion reference-based super-resolution method including flow-guided feature alignment and attention-guided feature adaptation, fused adaptively via confidence maps. On a public texture‑rich benchmark, this super-resolution method improves the peak signal-to-noise ratio by 1.07 dB. On two steel defect datasets, it achieved the best reconstruction quality in peak signal‑to‑noise ratio. When combined with an unmodified real-time detector, it improved the mean average precision by 6.9% and 1.2% over the best Transformer-based dedicated detection models on the two datasets, respectively. These findings confirm that high-fidelity super-resolution can compensate for the simplicity of the detector, validating its effectiveness in industrial defect detection applications.