Lightweight Image Super-Resolution Based on Re-Parameterization and Self-Calibrated Convolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing convolutional neural network methods for image super-resolution (SR) suffer from deep layers and numerous parameters, leading to feature loss and poor deployability on resource-constrained devices.
• Lightweight SR methods often overlook the importance of positional information, primarily relying on channel attention mechanisms.

Purpose of the Study:

• To propose a novel lightweight image super-resolution network (RepSCN) addressing the limitations of existing methods.
• To improve image clarity and reconstruction performance while reducing computational cost and model parameters.

Main Methods:

• Designed a re-parameterization distillation block (RepDB) and a self-calibrated distillation block (SCDB) to aggregate local distilled features from different receptive fields without extra parameters.
• Introduced a lightweight coordinate attention mechanism (CAM) to enhance feature representation at spatial and channel levels, incorporating crucial positional information.

Main Results:

• The proposed RepSCN network demonstrates superior reconstruction performance compared to classical lightweight super-resolution models.
• Achieved better image clarity and detail enhancement with significantly reduced model parameters and computational cost.

Conclusions:

• The RepSCN network offers an effective solution for lightweight image super-resolution, balancing performance and efficiency.
• The integration of re-parameterization, self-calibration, and coordinate attention mechanisms provides a promising direction for future SR research.