Image Motion Deblurring Based on Deep Residual Shrinkage and Generative Adversarial Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generative Adversarial Networks (GANs) for image deblurring often suffer from poor noise immunity and limited generalizability.
• Existing deblurring algorithms may require blur kernel estimation, adding complexity.

Purpose of the Study:

• To propose a novel network structure, DRSN-GAN, for robust and generalizable image motion deblurring.
• To enhance image quality, edge restoration, and texture details in deblurred images.

Main Methods:

• Developed a Deep Residual Shrinkage Network-Generative Adversarial Network (DRSN-GAN) for end-to-end image deblurring.
• Integrated a DRSN as the generator within a GAN framework to simultaneously denoise and deblur.
• Modified the DRSN architecture by repositioning Batch Normalization (BN) and ReLU layers before convolution for improved trainability.

Main Results:

• DRSN-GAN demonstrated superior subjective visual effects and objective evaluation metrics (PSNR, SSIM) compared to state-of-the-art methods like MPRNet.
• The model achieved improved restoration of image edges and textures, leading to higher overall image quality.
• YOLO detection accuracy was enhanced, and the number of parameters was reduced by 21.89%.

Conclusions:

• The proposed DRSN-GAN effectively addresses limitations of traditional GAN-based deblurring methods.
• This approach offers a robust solution for image motion deblurring with improved performance and efficiency.