Single-Channel Blind Image Separation Based on Transformer-Guided GAN.

Area of Science:

• Signal Processing
• Computer Vision
• Machine Learning

Background:

• Blind Source Separation (BSS) is challenging due to unknown source signal distributions and mixing matrices.
• Traditional BSS methods rely on statistical priors, which may not always apply.
• Existing Generative Adversarial Network (GAN)-based methods for blind image separation often neglect structural and detailed reconstruction, leading to residual interference.

Purpose of the Study:

• To propose a novel Transformer-guided GAN for improved blind image separation.
• To address the limitations of current GAN-based methods in reconstructing image structure and details.
• To enhance the accuracy and quality of separated images in BSS tasks.

Main Methods:

• Utilizing a Transformer-guided Generative Adversarial Network (GAN) with an attention mechanism.
• Employing adversarial training between a generator and a discriminator.
• Integrating a U-shaped Network (UNet) for convolutional feature fusion and structure reconstruction.
• Leveraging Transformer for positional attention to guide detailed information recovery.

Main Results:

• The proposed method demonstrates superior performance in blind image separation compared to existing algorithms.
• Quantitative experiments show significant improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
• The method effectively reconstructs image structure and details, minimizing residual interference.

Conclusions:

• The Transformer-guided GAN offers a robust solution for blind image separation, overcoming limitations of prior art.
• The integration of Transformer and UNet architectures enhances the reconstruction of structural and detailed information.
• This approach represents a significant advancement in the field of signal processing for image separation tasks.