An Accurate and Lightweight Method for Human Body Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Super-resolution (SR) is crucial for enhancing low-resolution images.
• Existing SR methods often struggle with complex details in human body images.
• Incorporating human body specific prior information can improve SR performance.

Purpose of the Study:

• To develop an efficient super-resolution method for low-resolution human body images.
• To leverage multi-scale features and human body prior for enhanced detail reconstruction.
• To improve the performance of human image analysis tasks on low-resolution data.

Main Methods:

• Proposed a lightweight multi-scale block (LMSB) for feature aggregation.
• Developed a framework with image reconstruction and prior estimation branches.
• Utilized human parsing maps and non-subsampled shearlet transform (NSST) sub-bands for human body prior representation.

Main Results:

• The proposed method achieves state-of-the-art performance on the HumanSR dataset.
• Requires approximately 8x fewer parameters compared to existing methods.
• Significantly enhances performance in human parsing and pose estimation tasks for low-resolution images.

Conclusions:

• The novel approach effectively super-resolves low-resolution human body images.
• The method demonstrates efficiency in terms of parameter count.
• Improved image quality leads to better performance in subsequent human image analysis applications.