Image Deblurring Using Multi-Stream Bottom-Top-Bottom Attention Network and Global Information-Based Fusion and...

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image deblurring is an ill-posed problem, often modeled by Gaussian blur.
• Traditional methods struggle with fine details and sharp edges.
• Deep learning methods show promise but require further refinement for detail reconstruction.

Purpose of the Study:

• To develop an effective end-to-end deep learning algorithm for non-blind image deblurring.
• To improve the restoration of fine details and reconstruction of sharp edges in blurred images.
• To enhance feature extraction and computational efficiency in deblurring networks.

Main Methods:

• A multi-stream bottom-top-bottom attention network (MBANet) with an encoder-decoder structure was designed.
• A coarse-to-fine multi-scale strategy was employed for image processing.
• A global information-based fusion and reconstruction network was utilized for refining deblurred images.

Main Results:

• The proposed MBANet effectively integrates low-level cues and high-level semantic information.
• Experiments on GoPro and REDS datasets demonstrated superior performance over traditional and state-of-the-art methods.
• Quantitative metrics (PSNR, SSIM) and human vision assessments confirmed the method's effectiveness.

Conclusions:

• The developed MBANet algorithm significantly advances non-blind image deblurring capabilities.
• The method excels in restoring image details and edge sharpness.
• It offers improved computational efficiency and robustness for practical applications.