An image deblurring method using improved U-Net model based on multilayer fusion and attention mechanism

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image deblurring in dynamic scenes is crucial.
• Current deep learning methods have limitations in feature interconnections and receptive fields.
• U-Net offers adaptable feature integration with fewer parameters and good accuracy.

Purpose of the Study:

• To propose an improved U-Net model for enhanced image deblurring in dynamic scenes.
• To address limitations of existing deep learning deblurring techniques.
• To improve feature extraction and detail recovery for clearer images.

Main Methods:

• Developed an improved U-Net model incorporating a multilayer feature fusion (MLFF) module for cross-layer feature integration.
• Introduced a dense multi-receptive field attention block (DMRFAB) for detailed feature extraction.
• Proposed a Frequency Reconstruction Loss Function (FRLF) using fast Fourier transform to minimize frequency differences.

Main Results:

• The proposed method demonstrated superior visual deblurring effects.
• Achieved a peak signal-to-noise ratio (PSNR) of 31.53 and a structural similarity index (SSIM) of 0.948 on the GoPro dataset.
• Attained a PSNR of 31.32 and an SSIM of 0.934 on the Real Blur dataset.

Conclusions:

• The improved U-Net model effectively enhances image deblurring performance.
• The MLFF and DMRFAB modules significantly boost feature extraction capabilities.
• The FRLF contributes to improved frequency reconstruction, resulting in higher quality deblurred images.