Appendix A: Moral approval report
The scanned copy of the formal moral approval report for the endoscopic video dataset is proven in Fig. 8.
The scanned copy of the formal moral approval report for the endoscopic video dataset.
Appendix B: Extra experimental particulars
This appendix gives further particulars on the scientific dataset, knowledge splitting technique, artificial degradation settings, coaching configurations, downstream VSR coaching settings, and t-SNE visualization settings used within the experiments.
Artificial dataset data
The data of the dataset utilized in “Evaluating simulated LR photos” and “VSR on the artificial dataset” sections is proven in Desk 9.
Knowledge splitting technique
The info splitting technique used within the experiments is summarized in Desk 10. We divided the 200 chosen HR video sequences into two non-overlapping subsets. The primary subset was used because the source-domain HR knowledge for degradation studying. The second subset was processed by totally different artificial degradation settings to assemble the target-domain LR knowledge for adversarial degradation studying. As well as, 20 additional video sequences had been reserved for testing. The cut up was carried out on the video-sequence stage to keep away from data leakage.
Artificial degradation settings
The detailed artificial degradation settings are reported in Desk 11. This artificial benchmark is designed to cowl a number of typical degradation components which will seem in endoscopic movies, together with interpolation degradation, isotropic Gaussian blur, direction-dependent anisotropic blur, movement blur, and noise. All degradation settings are carried out below the (occasions 4) upscaling issue.
UDLM coaching settings
The coaching settings of the proposed UDLM are summarized in Desk 12. Throughout coaching, every enter pattern consists of a video clip with 10 consecutive frames. We crop HR patches of measurement (256 occasions 256) from the unique HR frames. Underneath the (occasions 4) super-resolution setting, the corresponding LR patch measurement is (64 occasions 64). The mannequin is first educated with the low-frequency loss through the warm-up stage. After that, the adaptive knowledge loss is launched, and the adaptive kernel is periodically up to date in line with the present conduct of the degradation generator.
Loss weights
The loss weights used for UDLM coaching are listed in Desk 13. The low-frequency loss and adaptive knowledge loss share the identical data-loss weight. After the warm-up stage, the information constraint is switched from the low-frequency loss to the adaptive knowledge loss.
Coaching settings of downstream VSR fashions
The coaching settings of the downstream VSR fashions are summarized in Desk 14. For truthful comparability, all pseudo-paired knowledge generated by totally different degradation studying strategies are used to coach the downstream VSR fashions below the identical configuration.
Actual scientific dataset statistics
The statistics of the actual scientific endoscopic dataset utilized in “Efficiency analysis in real-world endoscopic situations” part are proven in Desk 15.
t-SNE visualization settings
The detailed settings of the t-SNE visualization experiment in actual endoscopic situations are summarized in Desk 16.
