Source-linked AI summary
Enhanced Invertible Encoding for Learned Image Compression
Yueqi Xie, Ka Leong Cheng, Qifeng Chen
TL;DR
Learned image compression has improved mainly through entropy modeling, while autoencoder transformations still leave information loss unresolved. The paper introduces an enhanced INN-based encoding network with channel squeezing and feature enhancement, and reports superior performance to learned methods and codecs, including VVC, especially on high-resolution datasets.
Problem
Existing learned compression methods mainly improve entropy models, while their autoencoder transformations do not explicitly solve information loss between image and latent feature spaces.
Method
The paper uses an enhanced invertible encoding network with INNs, an attentive channel squeeze layer, and a feature enhancement module for image compression.
Results
The method outperforms existing learned methods and compression standards, including VVC (VTM 12.1), on Kodak, CLIC, and Tecnick, especially the two high-resolution datasets.
Takeaways & Limitations
Highly invertible transformations preserve more detailed information and support stronger compression performance than existing learned methods and codecs within the evaluated datasets.
Abstract
from arXiv · showhide
Although deep learning based image compression methods have achieved promising progress these days, the performance of these methods still cannot match the latest compression standard Versatile Video Coding (VVC). Most of the recent developments focus on designing a more accurate and flexible entropy model that can better parameterize the distributions of the latent features. However, few efforts are devoted to structuring a better transformation between the image space and the latent feature space. In this paper, instead of employing previous autoencoder style networks to build this transformation, we propose an enhanced Invertible Encoding Network with invertible neural networks (INNs) to largely mitigate the information loss problem for better compression. Experimental results on the Kodak, CLIC, and Tecnick datasets show that our method outperforms the existing learned image compression methods and compression standards, including VVC (VTM 12.1), especially for high-resolution images. Our source code is available at https://github.com/xyq7/InvCompress.
1 INTRODUCTION
Learned image compression has increasingly improved entropy modeling, but commonly retains autoencoder transformations that can lose information. The paper proposes an enhanced invertible encoding network and reports gains over learned methods and codecs, especially on high-resolution datasets.
- Recent learned compression work mainly improves entropy models that parameterize latent-feature distributions more accurately and flexibly.
- Autoencoder-based transformations can select important information for reconstruction but do not explicitly prevent information loss during encoding.
- INNs offer strict invertibility for preserving information, but their use in compression is difficult because high-dimensional outputs are hard to quantize and compress.
- The proposed network uses an attentive channel squeeze layer to stabilize training and adjust feature dimensions, plus a feature enhancement module to improve nonlinear representation capacity.
- The method outperforms existing learned methods and traditional codecs, including VVC (VTM 12.1), particularly on the CLIC Professional Validation and Tecnick high-resolution datasets.
2 RELATED WORK
Traditional codecs use handcrafted transformations, while learned methods replace these rules with neural transformations and optimize rate-distortion objectives. Existing VAE-based methods are effective and stable but leave encoding information loss unresolved, motivating invertible approaches.
- Traditional methods: Traditional codecs typically combine handcrafted transformation, quantization, and entropy coding, using modules such as DCT and DWT.
- Learned methods: Learned compression methods use neural networks to learn nonlinear transformations between image space and latent feature space.
- Learned methods: RNN-based methods progressively encode residual information but rely on binary representation at each iteration and cannot directly optimize rate during training.
- Learned methods: VAE-based methods enable end-to-end optimization of estimated bit rates and reconstructed-image distortion despite nondifferentiable quantization.
- Learned methods: Later VAE-based improvements primarily strengthen entropy models through hyperpriors, context factors, attention, and more flexible likelihood parameterizations.
- Learned methods: Despite their performance and stability, VAE-based methods do not explicitly solve encoding information loss, leaving neglected information generally unrecoverable during decoding.
- Invertible neural networks: INNs provide bijective mappings, efficient inversion, and tractable Jacobians for explicitly computing posterior probabilities.
3 METHOD
The method replaces conventional analysis and synthesis transforms with an INN-based architecture, augmented to reduce feature dimensions and improve nonlinear capacity. It retains an entropy-modeling hyperprior while making the image-to-feature transformation highly invertible.
- 3.1 Background: A baseline transform-coding model encodes image x into latent y, quantizes it to ŷ, entropy-codes ŷ, and decodes ŷ into reconstructed image x̂.
- 3.1 Background: Learned compression minimizes a weighted rate-distortion objective, with rate estimated from latent entropy and distortion defined by MSE or 1 − MS-SSIM.
- 3.1 Background: Hyperprior models add side information z and use synthesis-transformed ẑ to estimate latent-feature distribution parameters.
- 3.2 Proposed Method: The proposed approach enhances the analysis and synthesis transforms with an invertible network, feature enhancement module, and attentive channel squeeze layer.
- 3.2 Proposed Method: The INN architecture uses downscaling and affine coupling layers to provide invertible analysis and synthesis transforms.
- 3.2 Proposed Method: Invertibility is guaranteed mathematically even when the internal feedforward functions are not themselves invertible.
- 3.2 Proposed Method: The feature enhancement module adds nonlinear representation before the INN in a residual manner because strictly invertible networks can have limited nonlinear capacity.
- 3.2 Proposed Method: The attentive channel squeeze layer reduces redundant channels, while its inverse replicates and reshapes quantized latent features for decoding.
4 EXPERIMENTS
Experiments evaluate the proposed compression method on Kodak, CLIC, and Tecnick using rate-distortion metrics, comparisons with learned models and codecs, qualitative reconstructions, and component analyses. The method outperforms competing approaches, including VVC on aggregate AUC, while its high-resolution results and ablations support the proposed design.
- Experimental Setup: Experiments use Kodak, CLIC, and Tecnick datasets, with CLIC and Tecnick containing predominantly high-resolution images.Kodak contains 24 images at 768×512, CLIC contains 41 higher-resolution images, and Tecnick contains 100 images at 1200 × 1200.
- Experimental Setup: Rate-distortion performance is evaluated using PSNR, MS-SSIM, and bits per pixel, with AUC reported as an aggregate comparison measure.RD curves compare coding efficiency, while AUC helps rank methods with similar performance.
- Experimental Setup: The proposed method is compared with state-of-the-art learned compression models and JPEG, JPEG2000, WebP, BPG, and VVC using the CompressAI evaluation platform.VVC results use the official VTM 12.1 test model.
- Rate-distortion Performance: On Kodak, the method slightly outperforms VVC (VTM 12.1) and substantially exceeds existing learned methods and traditional image compression standards.The comparison uses rate-distortion curves, with MS-SSIM transformed to −10 log10(1−MS-SSIM) for clearer visualization.
- Rate-distortion Performance: On CLIC and Tecnick, the MSE-optimized method outperforms all other evaluated approaches, indicating strong performance on high-resolution images.The authors describe the method as more robust and promising for compressing high-resolution images.
- Rate-distortion Performance: VVC and the proposed method achieve similar PSNR, but the proposed method has better aggregate AUC than VVC (VTM 12.1).AUC values are reported because the rate-distortion curves show similar PSNR performance.
- Qualitative Results: At approximately 0.125–0.145 bpp on Kodak samples, the MSE-optimized method performs well against VVC and better than other codecs, while MS-SSIM optimization preserves more structural details than traditional codecs.JPEG and JPEG2000 are evaluated at their lowest quality because they cannot reach the specified bpp levels.
- Analysis of Attentive Channel Squeeze: The attentive channel squeeze layer introduces minor deviation that is less than or comparable to quantization error in most cases.The analysis evaluates deviation and scaled deviation on Kodak across quality levels.
5 CONCLUSION
The enhanced Invertible Encoding Network uses INNs to make image compression more invertible and reduce information loss. Experiments report improved performance over learned methods and compression standards, especially on high-resolution datasets.
- INNs model image compression as an invertible process rather than using existing autoencoder-style transformations.
- The attentive channel squeeze layer stabilizes training and adjusts feature dimensions, while the feature enhancement module increases nonlinear transformation capacity.
- The highly invertible architecture is designed to largely mitigate information loss during image compression.
- The method outperforms state-of-the-art learned compression methods and existing standards, including VVC (VTM 12.1), especially on two high-resolution datasets.