A Novel Detail Enhancement Method for Industrial Radiography Based on Scattering Component Separation

Industrial digital radiography (DR) is a critical nondestructive evaluation technology used in safety-sensitive manufacturing industries. The high dynamic range of 16-bit DR images, while capturing subtle material discontinuities, poses a formidable challenge for defect analysis because of the presence of scattering-induced noise and low contrast defect signatures. Conventional enhancement techniques often fail to balance noise suppression with detail preservation, whereas deep learning methods require extensive annotated data and lack physical interpretability. To address these limitations, this study presents a novel physics-aware enhancement framework based on a radiation-matter interaction model. The core innovation involves estimating and separating the scattering component from the total detected intensity to recover the direct transmission signal. The proposed method employs a multistage architecture that sequentially performs attenuation estimation, scattering modelling, residual scattering removal, edge sharpening, and adaptive-detail contrast enhancement. Experiments were conducted on 60 industrial DR images obtained from weld inspections of ship plates, boilers, and oil pipelines. Quantitative evaluation using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Spatial Frequency (SF) improvement metrics demonstrated that the proposed method outperformed conventional techniques (global histogram equalization, contrast-limited adaptive histogram equalization), a transform-domain method (Discrete Wavelet Transform), and a convolutional neural network (CNN)-based model. The framework achieves the highest average PSNR (up to 26.96), SSIM (up to 0.76), and improved SF (up to 1203.18), as well as the lowest variances in these metrics. It also delivers substantial relative improvements over traditional methods (for example, PSNR gains over HE exceeding 90%) and consistent incremental gains over the CNN benchmark. This study contributes a physically interpretable, training-free solution that effectively enhances defect visibility and structural fidelity in industrial DR imagery, offering a reliable tool for practical nondestructive evaluation.