# Current-Aware Temporal Fusion with Input-Adaptive Heterogeneous Mixture-of-Experts for Video Deblurring

## TL;DR

This paper introduces a new video deblurring framework that improves image clarity by focusing on current frames and adapting to input features.

## Contribution

The paper proposes a novel Current-Aware Temporal Fusion framework and an input-adaptive Mixture-of-Experts module for efficient video deblurring.

## Key findings

- The proposed method achieves 33.09 dB PSNR and 0.9453 SSIM on the BSD dataset under severe blur conditions.
- The framework balances deblurring quality and runtime efficiency with minimal error accumulation.
- It supports both sequential and parallel inference, showing strong performance across multiple datasets.

## Abstract

In image sensing, measurements such as an object’s position or contour are typically obtained by analyzing digitized images. This method is widely used due to its simplicity. However, relative motion or inaccurate focus can cause motion and defocus blur, reducing measurement accuracy. Thus, video deblurring is essential. However, existing deep learning-based video deblurring methods struggle to balance high-quality deblurring, fast inference, and wide applicability. First, we propose a Current-Aware Temporal Fusion (CATF) framework, which focuses on the current frame in terms of both network architecture and modules. This reduces interference from unrelated features of neighboring frames and fully exploits current frame information, improving deblurring quality. Second, we introduce a Mixture-of-Experts module based on NAFBlocks (MoNAF), which adaptively selects expert structures according to the input features, reducing inference time. Third, we design a training strategy to support both sequential and temporally parallel inference. In sequential deblurring, we conduct experiments on the DVD, GoPro, and BSD datasets. Qualitative results show that our method effectively preserves image structures and fine details. Quantitative results further demonstrate that our method achieves clear advantages in terms of PSNR and SSIM. In particular, under the exposure setting of 3 ms–24 ms on the BSD dataset, our method achieves 33.09 dB PSNR and 0.9453 SSIM, indicating its effectiveness even in severely blurred scenarios. Meanwhile, our method achieves a good balance between deblurring quality and runtime efficiency. Moreover, the framework exhibits minimal error accumulation and performs effectively in temporal parallel computation. These results demonstrate that effective video deblurring serves as an important supporting technology for accurate image sensing.