A deep-supervised framework utilizing edge features and global texture features for distortion rectification in fisheye images

Fisheye cameras, with their ultra-wide field of view, are increasingly deployed in urban environments. However, the severe radial distortion inherent in fisheye imagery presents formidable challenges for downstream computer vision tasks, including object recognition, motion estimation, and semantic segmentation. To address this issue, a deeply supervised framework for distortion rectification that leverages edge features and global texture information (DR-EGT) is proposed, which uniquely integrates edge-aware structural features with global texture information under a unified deep supervision strategy. Unlike conventional approaches that rely solely on texture priors or geometric assumptions, DR-EGT introduces a hierarchical supervision mechanism that simultaneously leverages low-level edge contours and high-level texture semantics to guide the distortion correction process. This joint supervision enables the network to learn the distortion flow field of the entire image, which facilitates the reconstruction of geometrically accurate and perceptually sharp undistorted images. On the Places2 dataset, DR-EGT outperforms existing advanced methods, achieving a 12.12 % increase in PSNR (Peak Signal-to-Noise Ratio), 7.44 % improvement in SSIM (Structural Similarity Index Measure), 16.42 % reduction in MAE (Mean Absolute Error), and a 78.17 % gain in FID (Fréchet Inception Distance), demonstrating its superior reconstruction fidelity and perceptual quality. The results demonstrate the ability of DR-EGT not only correct complex fisheye distortion but also suppress post-correction artifacts and visual degradation.