DAN: Distortion-aware Network for fisheye image rectification using graph reasoning

Abstract

Despite the wide-field view of fisheye images, their application is still hindered by the presentation of distortions. Existing learning-based methods still suffer from artifacts and loss of details, especially at the image edges. To address this, we introduce the Distortion-aware Network (DAN), a novel deep network architecture for fisheye image rectification that leverages graph reasoning. Specifically, we employ the superior relational understanding capability of graph technology to associate distortion patterns in different regions, generating an accurate and globally consistent unwarping flow. Meanwhile, during the image reconstruction process, we utilize deformable convolution to construct same-resolution feature blocks and employ skip connections to supplement the detailed information. Additionally, we introduce a weight decay-based multi-scale loss function, enabling the model to focus more on accuracy at high-resolution layers while enhancing the model’s generalization ability. To address the lack of quantitative evaluation standards for real fisheye images, we propose a new metric called the “Line Preservation Metric.” Through qualitative and quantitative experiments on PLACE365, COCO2017 and real fisheye images, the proposed method proves to outperform existing methods in terms of performance and generalization.