Context-adaptive Entropy Model for End-to-end Optimized Image Compression
Authors: Jooyoung Lee, Seunghyun Cho, Seung-Kwon Beack
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Based on the experimental results, the proposed method outperforms the traditional image codecs, such as BPG and JPEG2000, as well as other previous artificial-neural-network (ANN) based approaches, in terms of the peak signal-to-noise ratio (PSNR) and multi-scale structural similarity (MS-SSIM) index. |
| Researcher Affiliation | Academia | Jooyoung Lee, Seunghyun Cho & Seung-Kwon Beack Broadcasting Media Research Laboratory Electronics and Telecommunications Research Institute Daejeon, Korea {leejy1003,shcho,skbeack}@etri.re.kr |
| Pseudocode | No | The paper describes the architecture and processes using text, equations, and diagrams (Figure 3, Figure 4, Figure 6), but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The test code is publicly available at https://github.com/Jooyoung Lee ETRI/CA_Entropy_Model. |
| Open Datasets | Yes | For training, we used 256 256 patches extracted from 32,420 randomly selected YFCC100m (Thomee et al. (2016)) images. |
| Dataset Splits | No | The paper states it used '256x256 patches extracted from 32,420 randomly selected YFCC100m (Thomee et al. (2016)) images' for training and '24 PNG images of the Kodak Photo CD image dataset (Kodak, 1993)' for evaluation, but does not specify a separate validation set split used during training. |
| Hardware Specification | Yes | The test was conducted under CPU environments, Intel i9-7900X. |
| Software Dependencies | No | Tensorflow and Python were used to setup the overall network structures, and for the actual entropy coding and decoding using the estimated model parameters, we implemented an arithmetic coder and decoder, for which the source code of the Reference arithmetic coding project2 was used as the base code. (No version numbers provided for TensorFlow or Python). |
| Experiment Setup | Yes | Each batch consists of eight images, and 1M iterations of the training steps were conducted, applying the ADAM optimizer (Kingma & Ba (2015)). We set the initial learning rate to 5e-5, and reduced the rate by half every 50,000 iterations for the last 200,000 iterations. Note that, in the case of the four λ configurations for high bpp, in which the hybrid entropy model is used, 1M iterations of pre-training steps were conducted using the learning rate of 1e-5. Although we previously indicated that the total loss is the sum of R and λD for a simple explanation, we tuned the balancing parameter λ in a similar way as Theis et al. (2017), as indicated in equation (6). We used the λ parameters ranging from 0.01 to 0.5. |