Cascaded Deep Video Deblurring Using Temporal Sharpness Prior and Non-Local Spatial-Temporal Similarity

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Video deblurring is challenging due to non-uniform blur across frames.
• Existing methods often lack efficiency and compactness.
• Leveraging inherent video properties can enhance deblurring performance.

Purpose of the Study:

• To develop compact and effective deep convolutional neural networks (CNNs) for video deblurring.
• To integrate domain knowledge of videos into CNN architectures for improved performance.
• To propose novel methods for exploiting temporal information and spatial-temporal similarities.

Main Methods:

• Developed a CNN integrating a temporal sharpness prior (TSP) to exploit non-uniform blur properties.
• Employed a cascaded training approach to handle motion field complexities.
• Proposed a non-local similarity mining approach using self-attention for frame restoration.

Main Results:

• The proposed CNN is 3x smaller in parameters compared to state-of-the-art methods.
• Achieved at least 1 dB higher performance in Peak Signal-to-Noise Ratios (PSNRs).
• Demonstrated superior performance on benchmark and real-world video datasets.

Conclusions:

• Exploring video domain knowledge leads to more compact and efficient CNNs for deblurring.
• The temporal sharpness prior and non-local similarity mining are effective strategies.
• The proposed method offers a favorable trade-off between model size and deblurring performance.