Visual-in-Visual: A Unified and Efficient Baseline for Image Restoration

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Image restoration research faces a trade-off between performance and computational efficiency.
• Existing methods often have limited applicability across various degradation types and datasets.

Purpose of the Study:

• To introduce VIVNet, a unified baseline model for image restoration that balances accuracy and efficiency.
• To demonstrate the model's versatility across a wide spectrum of image restoration tasks and datasets.

Main Methods:

• VIVNet integrates a biologically inspired micro visual module within a U-shaped architecture.
• The module employs lightweight depth-wise convolutions, similarity-aware weighting, and iterative element-wise multiplication.
• This design mimics human visual processing for enhanced feature extraction and dependency capture.

Main Results:

• VIVNet demonstrates competitive performance in image restoration tasks.
• The model achieves high computational efficiency, making it practical for real-world applications.
• Evaluations across general, all-in-one, composite degradation, UHD, underwater, medical, and remote sensing datasets confirm its robustness.

Conclusions:

• VIVNet offers a strong and efficient solution for image restoration, inspired by the human visual system.
• Its unified architecture and biologically inspired design enable high performance across diverse and challenging scenarios.
• The model presents a practical advancement in the field of image restoration.