Lattice Network for Lightweight Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep learning models, particularly those using residual blocks (RBs), excel in image restoration (IR).
• However, extensive use of stacked RBs leads to high memory and computational demands.
• Efficient IR models are crucial for practical applications.

Purpose of the Study:

• To design an economical and adaptive structure for connecting residual blocks in deep learning models for image restoration.
• To reduce the memory and computation costs associated with traditional residual block architectures.
• To enhance the representational capacity of image restoration models.

Main Methods:

• Proposed the lattice block (LB), inspired by lattice filters, utilizing butterfly-style topological structures to connect residual blocks.
• Employed adaptive channel reweighting to learn combination coefficients for LB structures.
• Integrated the LB as a plug-and-play module into various image restoration models.
• Introduced a novel contrastive loss for regularization to further improve model representation.

Main Results:

• Achieved significant reduction in model parameters (halved) and computational cost compared to traditional RB models.
• Maintained considerable reconstruction accuracy across various image restoration tasks.
• Demonstrated superior performance over state-of-the-art models in terms of accuracy, storage, and computation.
• The contrastive loss enhanced model representation without increasing inference expenses.

Conclusions:

• The proposed lattice block (LB) offers an efficient and effective alternative to conventional residual blocks for image restoration.
• LB facilitates the development of lightweight yet high-performance deep learning models for IR tasks.
• The method shows promise for advancing practical applications of image restoration techniques.