Bilateral Upsampling Network for Single Image Super-Resolution With Arbitrary Scaling Factors.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Single Image Super-Resolution (SISR) is crucial for various computer vision applications.
• Real-world scenarios often require SISR with arbitrary scaling factors due to fixed input/output image sizes.

Purpose of the Study:

• To address the challenge of developing a single model for SISR across arbitrary scaling factors.
• To propose an efficient and effective network for arbitrary scaling factor SISR.

Main Methods:

• A novel bilateral upsampling network is proposed, comprising a bilateral upsampling filter and a depthwise feature upsampling convolutional layer.
• The bilateral upsampling filter utilizes spatial and range upsampling filters for adaptive weight learning.
• The depthwise feature upsampling convolutional layer efficiently upsamples feature maps and recovers structural information.

Main Results:

• The proposed network demonstrates superior performance compared to state-of-the-art SISR methods on benchmark datasets.
• The adaptive nature of the bilateral upsampling filter allows for effective handling of varying scaling factors and pixel values.
• The depthwise convolutional layer efficiently reduces computational cost while preserving high-resolution feature details.

Conclusions:

• The bilateral upsampling network offers a robust solution for Single Image Super-Resolution with arbitrary scaling factors.
• The method achieves high-quality image reconstruction with enhanced detail recovery.
• This approach advances the capabilities of SISR in practical computer vision applications.