Image enhancement method for low-light pipeline weld X-ray radiographs based on weakly supervised deep learning

Abstract

X-ray inspection is the most intuitive approach for the non-destructive testing (NDT) of pipeline weld defects to avoid pipeline safety accidents. However, identifying pipeline weld defects in dark X-ray images is difficult due to low greyscale values. This paper proposed a weakly supervised network for denoising and enhancing low-light pipeline weld X-ray images. First, a semi-supervised network based on an improved Retinex-Net which implemented by self-paced learning was proposed to enhance illumination, yielding more natural X-ray images without artifacts, distortion, and overexposure. A new denoising network constrained by the X-ray images themselves was designed to achieve denoising while preserving the image detail. Qualitative comparison and quantitative analysis indicated that the proposed method outperformed other industrial image enhancement methods used for pipeline weld detection in terms of both subjective visual effects and objective metric values.