Frequency-Domain Optimized Laplacian Pyramid Network for Weld DR Image Enhancement

Welding quality is critical to the safe operation of key structures in aerospace, nuclear power, and related fields. Digital radiography (DR), widely used as a non-destructive testing method, plays a crucial role in weld defect detection. However, DR images often suffer from blurred details and low contrast due to various factors such as imaging conditions, equipment noise, and material properties, leading to insufficient defect visibility. To address this, this paper proposes a frequency-domain optimized Laplacian pyramid network (FDO-LPN). The proposed method effectively leverages the multi-scale decomposition capability of the Laplacian pyramid to finely enhance each frequency layer. Specifically, low-frequency components are enhanced using a nonlinear encoder-decoder module to highlight weld structures and morphological features. High-frequency components are processed by a local–global transformation module, which simultaneously captures local pixel correlations and long-range dependencies, thereby generating adaptive masks to optimize high-frequency details across various frequency bands. Subsequently, a multi-scale reconstruction module progressively refines high-frequency information. Finally, a trilinear regulation module globally adjusts brightness, contrast, and grayscale distribution, further enhancing overall image quality. Experimental results show that FDO-LPN achieves the best performance on objective evaluation metrics, including peak signal-to-noise ratio, structural similarity, and brightness difference, reaching 44.33, 0.99, and 0.46, respectively, with PSNR improved by 12.10 over the second-best method. In practical applications to high-resolution DR images of small-diameter pipe and flat-plate welds from different real DR systems, FDO-LPN notably expands the grayscale dynamic range, effectively restores structural information, and improves defect visibility across various weld types, thereby demonstrating strong practical applicability and generalization ability for engineering deployment.