Super-resolution of Fisheye Rectified Image based on Deep Multi-Path Cascaded Network

• Main Authors: Li, Weili, Li, Zhe, Wang, Haoming, Yin, Xiaoqing, Ma, YiQin
• Format: Conference Proceeding
• Language: English
• Subjects: Deep Learning; Distortion; Fisheye Rectified Image; Image reconstruction; Imaging; Network architecture; Reconstruction algorithms; Super-resolution; Superresolution; Training
• ISSN: 2771-6902
• DOI: 10.1109/BigDIA60676.2023.10429395

Fisheye imaging technique plays an important role in various areas. However, fisheye lenses can cause severe image distortion, which requires distortion rectification for these images. However, after distortion rectification, the images will suffer from the problem of resolution reduction, and the degree of resolution reduction varies in different regions. This problem will directly affect the subsequent process of image content analysis. Therefore, it is needed to adopt a super-resolution reconstruction algorithm to enhance the resolution of the fisheye rectified image. This paper designs a multi-path cascaded deep network which performs super-resolution reconstruction of fisheye rectified images. Through the multi-path deep network structure, the features of the fisheye rectified image are learned from multiple aspects at the same time, and the residual learning approach is introduced to further diminish the difficulty of network training and improve the super-resolution results. The mean square error loss is used to supervise the training process of the proposed deep work. The super-resolution reconstruction model is obtained which takes the input low-resolution fisheye rectified image as input and generate the super-resolution result. The results of super-resolution reconstruction experiments demonstrate that the proposed multi-path cascaded deep network can effectively enhance the resolution of fisheye rectified images, and can provide higher quality images for fisheye image analysis.