Reparameterizable large kernel attention networks for infrared image super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing infrared image super-resolution algorithms face challenges in balancing reconstruction performance and inference speed.
• Developing efficient algorithms is crucial for real-time infrared imaging applications.

Purpose of the Study:

• To propose a novel Large Kernel Reparameterization Attention mechanism for infrared image super-resolution.
• To develop a reparameterizable large kernel attention network that optimizes both performance and speed.

Main Methods:

• Introduced a Large Kernel Reparameterization Attention mechanism.
• Developed a reparameterizable large kernel attention network for infrared image super-resolution.
• Employed a multi-branch network during training and a single-branch network during inference for efficiency.

Main Results:

• Achieved a 0.0008 dB improvement in average PSNR on a self-constructed infrared dataset compared to state-of-the-art methods.
• Demonstrated fast inference speed, requiring only 37ms for 4× super-resolution on 320×180 images on an RK3588 Neural Processing Unit.

Conclusions:

• The proposed reparameterizable large kernel attention network effectively balances reconstruction performance and inference speed.
• This approach offers a significant advancement for infrared image super-resolution tasks, enabling real-time applications.