Image Deblurring by Exploring In-Depth Properties of Transformer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generative models have advanced image deblurring, but simultaneously improving perceptual quality and quantitative scores remains challenging.
• Existing methods struggle to balance visual fidelity with objective performance metrics in image restoration.

Purpose of the Study:

• To introduce pretrained transformers for image deblurring, aiming to enhance both perceptual quality and quantitative scores.
• To leverage deep features from pretrained vision transformers (ViT) for sharper image recovery.

Main Methods:

• Utilizing deep features from pretrained vision transformers (ViT) to capture global topological image relations and identify blur-induced changes.
• Developing two novel perceptual losses based on comparing transformer features (vector discrepancy and distribution comparison) between recovered and target images.
• Comparing transformer features to measure image sharpness and guide the deblurring process.

Main Results:

• Demonstrated improved perceptual quality in deblurred images without sacrificing quantitative scores like peak signal-to-noise ratio (PSNR).
• Outperformed competitive models (Uformer, Restormer, NAFNet) on defocus and motion deblurring tasks.
• Showcased the effectiveness of transformer properties in high-resolution image deblurring.

Conclusions:

• Pretrained transformers offer a powerful tool for image deblurring, effectively addressing the challenge of simultaneous perceptual and quantitative improvement.
• The proposed perceptual losses, guided by transformer features, provide a robust method for sharper image recovery.
• This approach advances the state-of-the-art in image deblurring, particularly for complex scenarios like defocus and motion blur.