基于径向先验与残差学习的畸变校正方法

Distortion Correction Method Based on Radial Prior and Residual Learning

摘要: 图像畸变是广角镜头、鱼眼镜头等大视场成像系统的固有缺陷，直接影响后续计算机视觉任务的精度。针对现有畸变校正方法物理先验缺失且难以应对透镜装配误差等非理想畸变的问题，文章提出一种基于径向先验与残差学习的畸变校正方法。该方法采用双分支解耦设计：径向物理分支利用物理模型约束，通过Brown-Conrady模型校正主要的桶形或枕形畸变，保证几何结构的整体正确性；残差修正分支利用数据驱动优势，通过U-Net架构预测高频非刚性位移场，精准补偿装配误差与非对称畸变。双分支通过位移场逐像素叠加与可微网格采样实现端到端联合优化。实验结果表明，本方法PSNR为24.49 dB、SSIM为0.9145，直线偏差为1.81 px，为视觉测量场景下的图像畸变校正提供了可靠的技术方案。

Abstract: Image distortion is an inherent defect of wide-field imaging systems, such as wide-angle and fisheye lenses, which directly impacts the accuracy of subsequent computer vision tasks. To address the limitations of existing distortion correction methods, which lack sufficient physical priors and struggle with non-ideal distortions such as lens assembly errors, this paper proposes a distortion correction method based on radial prior and residual learning. The approach employs a dual-branch decoupling design: the radial physical branch leverages physical constraints, using the Brown-Conrady model to correct primary barrel or pincushion distortions, thereby ensuring the overall geometric consistency of the structure; the residual correction branch capitalizes on data-driven advantages, employing a U-Net architecture to predict high-frequency non-rigid displacement fields, which accurately compensate for assembly errors and asymmetric distortions. The two branches are jointly optimized end-to-end through pixel-wise displacement field superposition and differentiable grid sampling. Experimental results show that the proposed method achieves a PSNR of 24.49 dB, an SSIM of 0.9145, and a line deviation of 1.81 px, providing a reliable technical solution for image distortion correction in visual measurement scenarios.