GLIC: General Format Learned Image Compression

Authors: MingSheng Zhou, MingMing Kong

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our method achieves state-of-the-art (SOTA) in the field of learned image compression in terms of PSNR and MS-SSIM. ... Comprehensive experiments are conducted in order to demonstrate the superiority of GLIC in the following aspects: the RD performance, the effectiveness of arbitrary format image compression applications, the reliability of progressive decoding, and the validity of subjective visual judgment. In addition, to verify the effectiveness of the proposed one-channel scheme, the AARB module, and the CCCT module, we perform ablation studies.
Researcher Affiliation Academia Ming Sheng Zhou, Ming Ming Kong* School of Computer and Software Engineering, Xi Hua University, Cheng Du, Si Chuan, China, 610039 EMAIL, EMAIL
Pseudocode No The paper describes the model architecture and modules in text and diagrams (Figure 8, Table 1, Table 2) but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/Du Bian Jun-007/GLICGeneral-Format-Learned-Image-Compression
Open Datasets Yes During model training, we use a subset of the Image Net dataset (Deng et al. 2009)... As Figure 9, we validated the BPP-PSNR and BPP-MSSSIM (d B) of the GLIC model in the Kodak (Eastman Kodak Company 1999)... For example, this method has undergone testing on specialized medical imaging datasets, specifically the subset0 dataset from Lung Nodule Analysis 2016 (LUNA16) (Setio et al. 2017), and the validation set from the Musculoskeletal Radiographs (MURA) (Rajpurkar et al. 2018)... An example of this involves the In-phase and Quadrature phase (I/Q) data from Bluetooth devices (Uzundurukan, Dalveren, and Kara 2020).
Dataset Splits Yes During model training, we use a subset of the Image Net dataset (Deng et al. 2009), including 14,206 images for training and 3,465 images for testing.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for the experiments.
Software Dependencies No The paper mentions 'Adam' as the optimizer but does not specify version numbers for any software libraries or frameworks used in the implementation (e.g., PyTorch version, Python version, CUDA version).
Experiment Setup Yes The optimizer we employ is Adam (Kingma and Ba 2014). The hyperparameter λ is set to these values: {0.002, 0.0042, 0.0075, 0.015, 0.03, 0.045, 0.07, 0.09}. Using a learning rate of 1 10 4 and batch size=16, the main network is trained for 100 epochs. Subsequently, 2000 epochs of fine-tuning with a learning rate of 1 10 5 ensure its convergence. After the above training, all parameters of the main network are frozen, only the gradient of g(<8) s is turned on, and independent training is performed using a learning rate of 1 10 5. Each preview synthesizer requires 2000 epochs,the lower the preview level,the more training time is required.