Early Exit or Not: Resource-Efficient Blind Quality Enhancement for Compressed Images (ECCV 2020)

Background

Official repository of Early Exit or Not: Resource-Efficient Blind Quality Enhancement for Compressed Images, ECCV 2020. [速览]

• A single blind enhancement model for HEVC/JPEG-compressed images with a wide range of Quantization Parameters (QPs) or Quality Factors (QFs).
• A multi-output dynamic network with early-exit mechanism for easy input.
• A Tchebichef-moments based NR-IQA approach for early-exit decision. This IQA approach is highly interpretable and sensitive to blocking energy detection.

Feel free to contact: ryanxingql@gmail.com.

Code & Pre-trained Model

We have released two versions of the RBQE approach.

• ECCV paper version

  • Adopts the RAISE data-set (high-quality and large-scale).
  • Adopts the HM software for compression and get YUV images.
  • Enhances only Y channel and report the Y-PSNR result (in accordance to previous papers).
  • Implements the image quality assessment module with MATLAB (convenient for visualization).
  • Assesses the Y quality; the IQA threshold is determined according to the Y performance.

• Improved version

  • Adopts the DIV2K data-set (used by most image restoration approaches).
  • Adopts the BPG software for compression (faster, but the result is different to that of the HM) and get PNG images (simpler).
  • Enhances RGB channels (more practical).
  • Re-implements the MATLAB-based image quality assessment module with Python (no need to use MATLAB anymore; but much, much slower than the MATLAB version).
  • Assesses the R quality; the IQA threshold is determined according to the R performance.

We have released the codes of all compared approaches in the latter repository.