Unifying Dimensions: A Linear Adaptive Mixer for Lightweight Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Super-Resolution

Background:

• Window-based Transformers excel in super-resolution via local self-attention (SA).
• However, Transformers suffer from high computational complexity and inference latency compared to convolutional neural networks (CNNs).
• Existing methods struggle to balance adaptive modeling with computational efficiency.

Purpose of the Study:

• To develop an efficient convolution-based Transformer for image super-resolution.
• To address the high computational complexity and latency of existing Transformer models.
• To enhance adaptive modeling capabilities while maintaining CNN-level efficiency.

Main Methods:

• Proposing a Focal Separable Attention (FSA) mechanism for linear computational complexity and long-range dynamic modeling.
• Introducing a dual-branch structure with an Information Exchange Module (IEM) for improved token mixing.
• Modifying feedforward networks into a Dual-Gated Feed-Forward Network (DGFN) to preserve high-dimensional channel information.

Main Results:

• The proposed Linear Adaptive Mixer Network (LAMNet) framework achieves superior performance in super-resolution tasks.
• LAMNet demonstrates significant improvements over existing Transformer-based methods.
• Achieved a 3x speedup in inference time compared to previous models, maintaining CNN computational efficiency.

Conclusions:

• LAMNet effectively combines the adaptive modeling of Transformers with the efficiency of CNNs.
• The proposed FSA, IEM, and DGFN modules contribute to enhanced performance and efficiency.
• LAMNet offers a promising direction for efficient and high-performance deep learning models in computer vision.