Video Deblurring with Adaptive High-frequency Extraction
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: video deblurring, deep learning, unsharp masking
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: State-of-the-art video deblurring methods use deep network architectures to recover sharpened video frames. Blurring especially degrades high-frequency information yet this aspect is often overlooked by recent models that focus more on enhancing architectural design. The recovery of high frequency detailing can be non-trivial, in part due to the spectral bias of neural networks. Neural networks are biased towards learning low frequency functions, making it to prioritize learning low frequency components. To enhance the learning of latent high frequencies, it is necessary to enforce explicit structures to capture the fine details or edges. This work merges the principles of the classic unsharp masking with a deep learning framework to emphasize the essential role of high-frequency information in deblurring. We generate an adaptive kernel, constructed from a convex combination of dynamic coefficients and predefined high-pass filtering kernels. This kernel is then employed in a spatio-temporal 3D convolution process to extract high-frequency components from the data. This method significantly improves video deblurring, achieving a noteworthy enhancement with an increase of up to 0.61dB in PSNR over top models on GORPO dataset. Additionally, it outpaces the majority of them in inference time.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3428
Loading