【摘要】 图像与视频编码及相应标准自诞生以来，一直支撑着点播、直播以及视频会议等核心多媒体服务。过去30余年，主流技术路线围绕规则驱动的模块化工具（如变换、预测、熵编码、环路滤波等）的精细化设计与协同优化展开，并借助标准化组织形成生态。近10年，随着深度学习表征能力、公共数据集累积以及高效训练/推理框架的成熟，端到端智能编码技术快速迭代，在若干测试集与应用场景中展现出超越传统标准的压缩性能。本报告围绕图像编码，第1部分概述端到端智能编码主流框架演化主线；第2部分阐述率失真性能指标之外的可实用性功能，包括可变码率与码率控制、模型量化与鲁棒性；第3部分总结智能编码纳入/影响标准化进程的努力与现状；第4部分探讨从智能图像编码到智能视频编码的进一步拓展。希望本文能够为研究者与工程实践者提供系统化的思考视角，促进智能图像视频编码方法在产业级场景中的有序落地。更多还原

【Abstract】 Over the past three decades, image and video coding technologies and their associated international standards have served as the foundational compression engines underpinning core internet-scale multimedia services, ranging from on-demand streaming and live broadcasting to real-time video conferencing and social media sharing. Traditional approaches have predominantly followed a rule-driven, modular paradigm, in which carefully engineered components—such as intra and inter prediction, block-based transforms(e. g., DCT and DWT), scalar quantization, entropy coding, and in-loop filtering—are jointly optimized under classical rate-distortion(R-D) theory. This methodology, refined through successive generations of standards, including JPEG, H. 26×, MPEG, and AVS, has achieved remarkable efficiency and interoperability through the coordinated efforts of standardization bodies. However, over the past decade, a paradigm shift has been catalyzed by the rapid advancement of deep learning, resulting in the emergence of end-to-end learned image and video compression. End-to-end trainable systems, empowered by expressive neural architectures, large-scale public datasets, and mature training ecosystems, have demonstrated R-D performance that consistently surpasses conventional codecs on benchmark datasets. These learned systems are primarily built upon variational autoencoders(VAEs), which replace the handcrafted rules of traditional pipelines with a unified differentiable framework. In this architecture, an encoder utilizes analysis transforms to map image data into compact latent representations, while a decoder applies synthesis transforms to reconstruct the image. Unlike linear transforms in traditional coding, these transforms leverage powerful neural networks, evolving from early convolutional neural networks(CNNs) to advanced architectures incorporating attention mechanisms, Transformers, and Mamba-based state-space models. A critical challenge in this framework lies in the nondifferentiable nature of quantization. Methods, such as additive uniform noise, are used during training to approximate quantization errors while maintaining differentiability to enable end-to-end optimization. Meanwhile, straight-through estimators are utilized to pass gradients directly through quantization layers. Although uniform quantization remains standard, recent advancements have explored vector quantization and non-uniform quantization strategies to further refine feature representation. The core of compression efficiency in these systems lies in entropy modeling, which estimates the probability distribution of latent variables to minimize the bitrate. This field has significantly evolved from early factorized models that assumed statistical independence among latents. The introduction of the hyperprior structure, which utilizes auxiliary latent variables to model the spatial distribution parameters of the primary latents, marked a significant milestone in capturing dependencies. Subsequent innovations introduced autoregressive modeling, which predicts current features based on causal contexts in spatial or channel dimensions, further enhancing probability estimation accuracy. Recently, hierarchical autoregressive models have been developed to capture global and local contexts in a coarse-to-fine manner, pushing the boundaries of feature compactness and coding efficiency. Furthermore, the optimization objectives have expanded beyond pixel-level fidelity metrics, such as MSE and MS-SSIM, to include perceptual metrics and adversarial losses, allowing for a trade-off between signal distortion and perceptual quality. Beyond theoretical performance, the transition of learned coding from academic exploration to industrial application requires addressing practical dimensions, such as variable rate control, hardware efficiency, and robustness. Researchers have developed mechanisms involving multi-scale decomposition and feature modulation, where quality factors or maps scale latent variables or intermediate features, to support variable bitrates within a single model. Rate control algorithms have also advanced, utilizing iterative search strategies or deep modeling of the rate-parameter relationship to meet specific bandwidth constraints. Model quantization techniques, including quantization-aware training and post-training quantization, are utilized to convert floating-point models into fixed-point integer operations, ensuring cross-platform consistency, reducing computational overhead, and facilitating deployment on commodity hardware. Furthermore, robust coding frameworks are designed to defend against adversarial attacks and transmission errors through training regularization and input preprocessing, addressing the vulnerability of neural networks to perturbations. These technical advancements have culminated in the integration of learned compression into formal international standards. Two landmark standards have recently emerged: JPEG AI, ratified by ITU-T as T. 840. 1, and IEEE 1857. 11-2024. Although both standards adopt VAE-based architectures, they differ in design philosophy. JPEG AI utilizes a multibranch network design and emphasizes subjective quality and machine-task compatibility, optimizing for perceptionoriented metrics, such as MS-SSIM and VMAF. By contrast, IEEE 1857. 11 focuses on objective gains in PSNR and MSSSIM, offering tiered complexity profiles(base, main, and high) to adapt to different computational capabilities. Both standards have established rigorous training and evaluation protocols, including the use of specific datasets, such as Kodak and dedicated robustness benchmarks, to ensure a fair comparison and reproducibility. The principles of learned image coding have naturally extended to video coding, although with unique challenges in temporal modeling. The evolution of neural video coding can be categorized into three developmental phases. The first phase involved hybrid approaches that replaced specific modules, such as intra prediction, with learned networks while retaining the traditional motioncompensated residual coding framework. The second phase progressed toward conditional inter-frame coding, utilizing learned optical flow networks for motion estimation and warping to generate temporal contexts. The third and most recent phase marks a shift toward unified probabilistic frameworks that entirely eliminate explicit motion estimation. These systems leverage hierarchical spatial-temporal priors to perform joint intra and inter prediction within a single model, achieving performance that rivals or exceeds the latest H. 266/VVC standard while approaching real-time processing speeds on GPUs. Future directions indicate that the field is converging toward two major trends: task-aware coding and generative integration. Task-aware coding aims to support human vision and machine perception from a single bitstream, aligning with the biological principle of “compression as intelligence”, where compact representations facilitate diverse downstream cognitive tasks. Furthermore, the integration of generative models, such as diffusion and large multimodal models, is enabling ultra-low-bitrate reconstruction with high semantic fidelity, fundamentally altering the rate-distortion-perception trade-off. This report synthesizes these technical, practical, and standardization advances to provide a comprehensive perspective. Finally, the future of intelligent compression lies in establishing a new foundation for multimodal, task-agnostic, and semantically aware visual communication.更多还原