Method on fisheye image rectification based on distortion feature extraction

Fisheye image distortion rectification is a challenging task. Existing fisheye image rectification methods struggle to balance rectification accuracy and computational efficiency. To address this issue, this paper proposes a novel fisheye image rectification method based on distortion feature extraction. First, a Transformer UNet with Cross-Shaped Windows (CSWin-UNet) is employed, incorporating progressive down-sampling and a Convolution Block Attention Module (CBAM) at the input stage to mitigate information loss during down-sampling while enhancing local distortion feature extraction. Subsequently, the Patch Merging structure is integrated into the CSWin-UNet encoder to reduce model complexity and improve feature integration. Finally, a progressive training method using distortion distribution maps and pixel-level flow maps as supervisory signals enables the model to focus on distortion feature extraction, ultimately generating pixel-level flow maps for fisheye image rectification. Experiments demonstrate that the proposed deep learning-based method achieves 26.799 dB PSNR and 0.921 SSIM with only 6.06 GFLOPs on Place365- synthesized fisheye images, enabling efficient and accurate correction.