Review — Han VCIP’20: HDR Image Compression with Convolutional Autoencoder (HDR JPEG Image Compression)

HDR Images Are Extremely Useful When There are Places That Are Bright and Dark Within the Same Image

In this paper, HDR Image Compression with Convolutional Autoencoder, (Han VCIP’20), is briefly reviewed. In this paper:

  • A two-layer High Dynamic Range (HDR) image compression framework based on convolutional neural networks is proposed.

This is a paper in 2020 VCIP. (Sik-Ho Tsang @ Medium)


  1. Proposed Two-Layer HDR Image Compression Framework

1. Proposed Two-Layer HDR Image Compression Framework

1.1. Encoder Framework

HDR Encoder Framework
  • The original HDR image is tone-mapped (Tone-Mapping Operation, TMO) to obtain the Low Dynamic Range (LDR) image, and encoded by JPEG encoder.
  • At the extension layer, a CNN autoencoder is used to encode the residual image.

1.2. Decoder Framework

HDR Decoder Framework
  • First, the base layer codestream is decoded by JPEG decoder and obtained the LDR image.

2. Proposed Autoencoder & Post-Processing Networks

Proposed Autoencoder
  • Residual decoder and encoder are composed of convolution, deconvolution and GDN/IGDN.

2.1. Residual Encoder & Decoder

  • In the residual encoder, a feature fusion structure to concatenate and fuse features of different convolutional layers.

2.2. Binarizer

  • It is used to control and reduce the length of coded stream.
  • But the above equation is not good for back-propagation.
  • where ε is random noise.

The B(yijk) function is only used in forward propagation calculation, while ~B(yijk) is used in back propagation.

  • The gradient of ~B(yijk) function can be obtained by:

2.3. Residual Reconstruction And Rate Control

Residual reconstruction method based on iteration and accumulation

An iterative accumulation way to control the extension layer bit rate and reconstruct the original residual image.

  • In the first iteration, the input image of the encoder is the original residual image.

2.4. Post-Processing

Post-processing architecture
  • The post-processing module is a two-layer group convolutional neural networks. The network is mainly composed of 3×3 and 5×5 residual blocks.

3. Experimental Results

3.1. Training Set

  • The original HDR image comes from the public Internet HDR image dataset and video sequence[15] (including HDReye, Fairchild, Funt, MPI, etc.).

3.2. Objective Quality

Objective image quality comparison
  • The HDR-VDP-2 index is adopted as the objective evaluation metric.

3.3. Subjective Quality

Subjective image quality comparison
  • For the red box which is the bright area, and the blue box which is the dark area, the subjective image quality of the proposed method is slightly better than JPEG XT profile A, B, C.

PhD, Researcher. I share what I've learnt and done. :) My LinkedIn:, My Paper Reading List:

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store