Adversarial Dual-Distortion Image Rectification: Integrating Multi-Scale Discriminators and Central Self-Attention Modules

Fisheye cameras, leveraging their wide-angle imaging advantages, have been widely adopted in industrial pipeline inspection, particularly for monitoring the inner wall conditions of cylindrical pipelines such as those in oil and gas wells. To achieve geometric reconstruction of pipeline wall images captured by fisheye lenses, this study proposes a method aimed at simultaneously correcting pipeline wall deformations (caused by cylindrical geometry) and fisheye lens distortions. Literature research reveals that fisheye distortion correction for pipeline wall images remains in its infancy, lacking real-scene datasets and data synthesis methods to support deep learning model training. Addressing this bottleneck, we constructed a dual-distortion mathematical model based on geometric optics principles and proposed a parametric simulation-based synthesis method to generate a specialized synthetic dataset tailored for training fisheye distortion correction networks. Existing fisheye correction methods exhibit performance degradation when processing dual-distortion images. To overcome this, we innovatively designed a central self-attention module adapted to fisheye imaging characteristics. This module dynamically enhances geometric feature extraction in the central region of distorted images, effectively suppressing distortion diffusion effects in peripheral areas. Additionally, we introduced a multi-scale discriminator with a pyramidal feature fusion architecture to strengthen the synergistic optimization of local textures and global structures in generative adversarial networks. Experimental results demonstrate that, compared to state-of-the-art correction methods, our approach achieves significant improvements in both objective quantitative metrics (e.g., SSIM, PSNR) and visual consistency. These findings validate the method’s robust capability to effectively correct fisheye-distorted pipeline wall images for industrial inspection applications, providing a reliable solution for related fields.