An Efficient Multiscale Spatial Rearrangement MLP Architecture for Image Restoration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Local window-based models offer linear complexity for high-resolution images but struggle with long-range context due to limited receptive fields.
• Efficiently encoding long-range information is crucial for advancing network performance in image restoration tasks.

Purpose of the Study:

• To develop a novel architecture for image restoration that effectively utilizes long-range information.
• To enhance the receptive field of local window-based models without increasing computational complexity.
• To achieve superior image restoration quality with reduced computational cost.

Main Methods:

• Proposes a single-stage multiscale spatial rearrangement multilayer perceptron (MSSR-MLP) architecture.
• Introduces a spatial rearrangement module (SRM) to integrate information from outside the local window.
• Employs multiple SRMs with varying step sizes to capture multiscale information.

Main Results:

• The MSSR-MLP effectively models long-range dependencies using window-based fully connected layers.
• Achieves improved image restoration performance across various tasks, including denoising, dehazing, and deblurring.
• Demonstrates significant reductions in computational cost (FLOPs) compared to existing state-of-the-art methods like SwinIR, C2 PNet, and MAXIM.

Conclusions:

• The proposed MSSR-MLP architecture offers an efficient and effective solution for image restoration.
• Spatial rearrangement is a viable technique for extending local receptive fields and capturing multiscale context.
• The method achieves a strong balance between performance and computational efficiency in image restoration.