The Booth algorithm is widely used for efficient signed multiplication due to its ability to reduce partial products. A higher radix Booth multiplier generates fewer partial products, while it also increases hardware complexity in the generator, diminishing the advantage of fewer accumulators. Previous optimizations of generators and accumulators were designed for application-specific integrated circuits (ASICs), but their performance gains cannot be comparably translated to field-programmable gate arrays (FPGAs) due to differences in architecture. This article proposes FPGA-friendly approximate Booth multipliers that combine approximate hybrid-radix partial product generation with resource-efficient accumulation techniques. Initially, to improve generation efficiency, an look-up table (LUT)-reused exact radix-8 generator is introduced through logical partitioning to integrate two types of partial products into a single LUT. In addition, approximate adjacent-compensation radix-8 and radix-16 generators are developed based on the Booth encoding bit-repetition principle. Later, to speed up partial product accumulation, an overlap-parallel accumulation scheme and various accumulators are proposed, reducing compression steps and enhancing resource utilization. Last, performance-configurable hybrid radix-8/-16 approximate Booth multipliers are designed to meet the needs of different error-resilient applications. The most hardware-efficient configuration of the proposed 16-bit multiplier reduces power-delay product (PDP) and LUT consumption by 38.31% and 35.66%, respectively, compared with the exact multiplier. Furthermore, the proposed designs offer a better balance between accuracy and hardware complexity than existing approximate multipliers. The practicality of these multipliers is demonstrated in both joint photographic experts group (JPEG) image compression and finite impulse response (FIR) filtering applications. An open-source library of the proposed multipliers is available at https://github.com/YnuGuoLab/FPGA_Signed_Approx_Mul to support further research.

With the rapid growth of online media, the number of images on the internet has surged, posing significant challenges for data transmission and storage. This necessitates the development of efficient compression algorithms. While learned image compression (LIC) methods have made significant progress, they often overlook the viewer's subjective experience. Unlike rate-distortion-oriented codecs, perceptual image compression (PIC) prioritizes preserving human-perceived visual quality. However, many existing PIC methods attend either to global context or to local details, failing to couple the two. We therefore present a PIC network that fuses multi-modal information across multiple scales, jointly modeling global semantics and fine-grained structures to improve perceptual fidelity and achieve competitive perceptual compression quality. We propose a fusion method called Multi-scale Multi-modal Merge Network (M$^{3}$Net) to effectively combine information from different modalities. Our approach first fuses multi-modal information along the height dimension to capture relevant vertical context, then reduces dimensionality and collects horizontal context along the width dimension. By alternating these operations, the final pixels effectively capture the overall context of the image. This method significantly improves the perceived quality of images while reducing model size and runtime. We conduct systematic ablations of the core components and perform large-scale evaluation on eight objective metrics spanning state-of-the-art perceptual quality, structural distortion, and traditional measures. Results show that PIC-M$^{3}$ achieves competitive or superior results on key perceptual metrics, providing multi-faceted evidence of its effectiveness. A human subjective preference study on the Kodak dataset further shows that the perceptual improvements of PIC-M$^{3}$ are consistent with human judgments.

Learned image compression methods have shown impressive performance but are often highly specialized for either human perception or specific machine vision tasks. This specialization limits their versatility and requires costly retraining for new applications. To address this, we introduce UniCodec, a universal codec built on a novel paradigm of semantic disentanglement at the encoder and compositional generation at the decoder. This framework is designed to simultaneously serve both human and machine needs, eliminating the need for task-specific retraining. At the encoder, UniCodec leverages pre-generated, task-specific label codebooks created by a Large Language Model (LLM). For any given task, a grounding model uses the corresponding codebook to perform task-aware disentanglement, compressing only the most relevant image regions. This mechanism not only saves significant bits but is also the key to our system’s rapid, zero-retraining adaptation: switching to a new task is as simple as selecting a new codebook. The decoder then performs compositional generation: it combines the compact, disentangled components with powerful priors from a generative diffusion model. This process reconstructs a high-quality, complete image optimized with rich detail for human perception and precise features for machine vision tasks. Extensive experiments demonstrate that UniCodec consistently outperforms existing methods, effectively bridging the gap between human-centric and machine-centric compression.

Video Coding for Machines (VCM) is an emerging topic aiming to bridge compression for human and machine vision tasks. While existing VCM methods excel at spatial tasks (e.g., object detection), they have largely neglected temporal-dependent tasks (e.g., action recognition). Furthermore, the common cascaded pipeline connecting compression and analysis networks induces information bottlenecks and wasteful computation. This paper proposes a VCM framework that addresses these issues by exploiting information entirely within the compressed domain. We introduce a Spatial-Temporal Decoupled Latent Composition module (STDLC) that intercepts and processes features directly from the compression pipeline. The spatial path is composed of Swin-transformer blocks, and in the temporal path, a novel Temporal-Swin (T-Swin) block, finalized with a strong latent composition. Our framework demonstrates superior post-compression action recognition accuracy compared to traditional and prior VCM approaches on 2 standard datasets, with reduced computational costs while maintaining perceptual quality.

In recent years, point clouds have received a lot of focus. This kind of data is applied to various 3D machine vision tasks, such as 3D object detection and 3D semantic segmentation, and further to 3D scene understanding. However, transmitting point clouds is costly. Therefore, some recent neural-network-based point cloud compression (PCC) methods have been proposed to reduce the transmission cost of point clouds. Beyond PCC, applying pre- and post-processing methods can further improve transmission efficiency while keeping performance on 3D machine vision tasks. In this paper, we propose a Region-of-interest (ROI) and Upsampling-enhanced point cloud Transmission (RUT) method for 3D machine vision. RUT reduces the point cloud transmission cost by retaining only the foreground (ROI) areas, while maintaining machine vision performance by applying point cloud upsampling (PU) after transmission. The experimental results show that, with the same transmission cost, our method achieves at least a 14.24 mAP@0.25mIoU improvement on 3D object detection and a 5.63 mIoU improvement on 3D semantic segmentation tasks compared with models without our RUT on the ScanNetV2 dataset.

Overfitted neural image codecs offer a promising path to ultra-low-complexity decoding, ideal for resource-constrained hardware. However, realizing their potential requires an integer-only pipeline, a task complicated by their extreme sensitivity to post-training quantization (PTQ). To our knowledge, this is the first work to systematically address the challenges of 8-bit PTQ for such codecs. We present an effective methodology for 8-bit weight, 16-bit activation quantization, developed through a three-stage investigation. (1) We first establish that a fine-grained, per-channel quantization scheme is the essential foundation for preserving model fidelity. (2) We then use Hessian analysis to identify strong weight dependencies as the primary bottleneck and apply a dependency-aware compensation algorithm to mitigate the resulting errors. (3) To further refine the compensation process, we integrate a principled, decreasing-importance order into the quantization sequence. Our final approach reduces the rate-distortion overhead to just +2.95% BD-Rate on the Kodak dataset and +2.78% BD-Rate on the CLIC dataset, establishing a practical methodology for this sensitive codec class.

Neural Video Codecs (NVCs) offer unprecedented rate-distortion performance, making them highly attractive for bandwidth-constrained environments like 5G cellular networks and emerging satellite direct-to-cell (D2C) links. However, deploying NVCs in real-world streaming applications is severely hindered by cross-platform floating-point non-determinism, which causes arithmetic entropy coders to desynchronize and crash across different GPU architectures. While recent integer-based quantization methods address this, they incur either massive degradation in compression efficiency (INT8) or severe computational bottlenecks by bypassing hardware acceleration (INT16). In this paper, we propose a streamable, client-server NVC architecture featuring a novel Mixed Precision (FP16/FP32) strategy. By strategically executing P-frames in hardware-accelerated FP16 for real-time throughput, while forcing I-frames and periodic feature-adapter resets to IEEE-754 compliant FP32, we guarantee deterministic synchronization at critical boundaries. Through extensive cross-encode/decode evaluations across 12 GPUs spanning four architectural generations, we demonstrate that our approach successfully eliminates intra-generation fragmentation and substantially broadens cross-die interoperability, achieving seamless cross-generation decodability for recent architectures at 1080p. Crucially, this is achieved with a negligible impact on compression efficiency. Furthermore, we evaluate the system's end-to-end latency across diverse real-world networks, including Wi-Fi 6, 5G NR (FDD/TDD), and Starlink D2C, proving the practical viability of streamable learned video compression while highlighting unique challenges in Non-Terrestrial Networks.

Learned image compression (LIC) models have achieved remarkable performance but typically rely on computationally expensive backbones, which limits their applicability on resource-constrained embedded devices. To address this limitation, we present GhostIC, a Ghost-driven lightweight image compression network that achieves efficient inference while maintaining competitive rate–distortion performance. In GhostIC, we design a GhostBlock inspired by GhostNet where redundant features are generated from intrinsic features using inexpensive operations. This design integrates prior knowledge of image spatial redundancy, enabling efficient representation for image compression. Furthermore, the GhostBlock integrates the decoupled fully connected (DFC) attention mechanism, which enhances long-range dependency modeling with minimal overhead and compensates for the representational limitations of lightweight convolutions. Extensive experiments on standard benchmarks demonstrate that GhostIC delivers a rate-distortion performance on par with VVC intra coding while substantially reducing computational complexity compared to state-of-the-art learned models. With its favorable balance between efficiency and performance, GhostIC provides a practical solution for real-time learned image compression on resource-constrained devices. The source code will be made available upon publication.

The rate-distortion-perception (RDP) trade-off extends classical rate–distortion theory by imposing a distributional constraint on reconstructions, providing a unified framework for neural image compression that jointly governs fidelity and perceptual realism. While prior work achieves near-optimal rate–perception trade-offs, practical frameworks explicitly realizing the full RDP surface remain scarce, primarily due to the difficulty of introducing common randomness at the decoder. We propose DCIC (Dual-Constrained Diffusion Image Compression), which integrates a learned codec with a diffusion-based decoder governed by joint distortion and idempotence constraints. The distortion constraint bounds reconstruction fidelity relative to the base codec output; the idempotence constraint – requiring that re-encoding the restored image recovers the base codec reconstruction – serves as a tractable surrogate for the distributional perception requirement. Together, they steer the reverse denoising process via iterative optimization with consistent noise injection, realizing common randomness without additional rate overhead. At fixed rate, dual attenuation factors (K_D, K_P) jointly navigate the Pareto frontier of the distortion-perception plane, enabling continuously adjustable fidelity-realism trade-offs from a single bitstream. DCIC_RD (K_P=0) and DCIC_RP (K_D=0) arise as boundary curves, with DCIC_RDP (K_D = K_P=1) realizing the optimal interior operating point. Experiments on CelebA-HQ, CLIC2020, and ImageNet-1K across CNN, Transformer, and hybrid architectures confirm that DCIC_RDP achieves superior BD-PSNR over all perceptual codecs, while DCIC_RP matches dedicated perception-oriented methods in BD-FID, validating the practical value of full RDP surface navigation.

Learned image compression (LIC) holds great potential for advancing various computer vision applications. However, traditional fixed-rate LIC methods require training separate models for each target bitrate, resulting in high computational costs and limited scalability. To address these challenges, we propose a parameter-efficient variable-rate compression framework that combines BitFit and Low-Rank Adaptation (LoRA). Specifically, we fine-tune only the bias terms and apply LoRA to the fully connected layers within the transformer blocks, while keeping all other parameters frozen. A pretrained or newly trained fixed-rate model serves as the backbone, which is adapted to other target rates through lightweight fine-tuning. Furthermore, we introduce trainable channel-wise quantization step sizes to enhance compression performance across different bitrates. Experimental results demonstrate that our method achieves competitive rate-distortion performance with significantly reduced training time and memory consumption, offering a practical solution for scalable variable-rate image compression.

Recent advances in learned image compression (LIC) have demonstrated superior performance over traditional methods but often require training and storage of multiple models to handle different bitrate settings. In this paper, we propose the Uniform Spatial-Frequency Residual Bottleneck Modulation Adapter (U-SFRB), a plug-and-play, adapter-based framework for variable rate image compression that significantly reduces training and storage overhead. Our method freezes the backbone network and only trains lightweight adapters—Spatial-Frequency Residual Bottleneck Adapters (SFRBs)—to achieve rate adaptability. By inserting multiple SFRBs in parallel, our approach enables a single model to support a wide range of bitrates. Unlike prompt-based methods restricted to transformer architectures, our approach is compatible with both CNN- and transformer-based compression models. Experimental results on the Kodak and CLIC datasets show that our method achieves competitive rate-distortion performance compared to state-of-the-art variable rate compression approaches, with the advantage of lower training complexity and better model flexibility.

The goal of this paper is to raise awareness of challenges and opportunities in the Collaborative Intelligence (CI) field and promote research on related standards. We begin by identifying a key challenge in CI applications, i.e., is it still possible for feature streams received in the cloud to be reused in the future by more advanced multitasking networks to achieve effective task accuracy? We then propose a framework to explore the generalization ability of cloud-received feature streams on more advanced networks from a coarse-grained to a fine-grained manner. We design a series of adapters of varying complexity to further explore the potential of feature streams for task network adaptation. Experiments show that sharing feature streams across multiple task networks could achieve an average of nearly 80% bitrate saving compared to Versatile Video Coding (VVC), which demonstrates the reuse potential of cloud-received feature streams. In addition, we make theoretical inferences about the adaptation range of shared feature streams, especially for those networks with high precision.

Learned image compression (LIC) has become more and more important in recent years. The hyperprior-module-based LIC models, which use hyperprior module to predict the distribution of image features and improve entropy coder performance, have achieved remarkable rate-distortion (RD) performance. However, the storage and memory costs of these LIC models are too high, resulting in higher difficulty to be applied to various devices, especially portable or edge devices. The storage and memory cost are directly linked to the parameter number. As a preliminary experiment, we manually assigned half channels for the hyperprior module in LIC models, reducing about 30% parameters in the model. The pruned models still kept similar RD performance to the original ones. This reveals that the hyperprior module in LIC models is highly redundant. In the meanwhile, LIC models with different reconstruction qualities require different amounts of parameters for the hyperprior module. Based on these phenomena, we propose a quality-aware hyperprior pruning method that efficiently reduces the storage and memory cost of the hyperprior module and various context models. It consists of two parts. The first part is the pruning method itself, called enhanced ResRep on hyper path (ERHP). The second part is a quality-aware threshold searching method, called pruning threshold searching (PTS), which prunes the hyperprior module based on the reconstruction qualities of LIC models. The experiments on various LIC models show that our methods reduce large volumes of storage cost (up to 74.6%) and memory cost (up to 41.5%), while keeping the performance the same before pruning.

In recent years, Learned Image Compression (LIC) has drawn significant attention because of its powerful coding ability. However, similar to most neural network-based schemes, LIC is vulnerable to adversarial attacks. To mitigate the influence of adversarial attacks, adversarial training is usually adopted to finetune the network. This paper proposes efficient adversarial attack and training methods for LIC, by proposing three losses based on the original image, adversarial image, and reconstructed image. For the attack, we study the effects of three proposed losses on four qualities of classical factorized-prior and hyperprior models. For the adversarial training, all the proposed three losses are used in the finetuning for the three attack scenarios, respectively. We find that using the loss between the adversarial output and the original image achieves strong defense performance against various attacks, improving RD cost by up to 68.6%. Furthermore, we show that updating only the decoding during adversarial training along with reducing the number of iterations, can reduce the training time by up to 82.9% without compromising the defense performance.

Learned image compression (LIC) has reached a comparable coding gain with traditional hand-crafted methods such as VVC intra. However, the large network complexity prohibits the usage of LIC on resource-limited embedded systems. Network quantization is an efficient way to reduce the network burden. This paper presents a quantized LIC (QLIC) by channel splitting. First, we explore that the influence of quantization error to the reconstruction error is different for various channels. Second, we split the channels whose quantization has larger influence to the reconstruction error. After the splitting, the dynamic range of channels is reduced so that the quantization error can be reduced. Finally, we prune several channels to keep the number of overall channels as origin. By using the proposal, in the case of 8-bit quantization for weight and activation of both main and hyper path, we can reduce the BD-rate by 0.61%-4.74% compared with the previous QLIC. Besides, we can reach better coding gain compared with the state-of-the-art network quantization method when quantizing MS-SSIM models. Moreover, our proposal can be combined with other network quantization methods to further improve the coding gain. The moderate coding loss caused by the quantization validates the feasibility of the hardware implementation for QLIC in the future.

LiDAR point cloud plays an important role in various real-world areas. It is usually generated as sequences by LiDAR on moving vehicles. Regarding the large data size of LiDAR point clouds, Dynamic Point Cloud Compression (DPCC) methods are developed to reduce transmission and storage data costs. However, most existing DPCC methods neglect the intrinsic misalignment in LiDAR point cloud sequences, limiting the rate–distortion (RD) performance. This paper proposes a Misalignment-aware Dynamic LiDAR Point Cloud Compression method (MDLPCC), which alleviates the misalignment problem in both macroscope and microscope. MDLPCC exploits a global transformation (GlobTrans) method to eliminate the macroscopic misalignment problem, which is the obvious gap between two continuous point cloud frames. MDLPCC also uses a spatial–temporal mixed structure to alleviate the microscopic misalignment, which still exists in the detailed parts of two point clouds after GlobTrans. The experiments on our MDLPCC show superior performance over existing point cloud compression methods.

This paper describes our design for learned image codec (LIC) on FPGA, from the aspects of algorithm, architecture and system. For the algorithm, we build the neural network on the hyperprior structure. Besides, we present a quantization aware training scheme specifically adapted to LIC. For the architecture, we propose a fine-grained pipeline architecture. Channel parallelism constraint and neural network search are proposed to improve the DSP utilization and efficiency, respectively. For the system, we make a CPU-FPGA heterogeneous coding system in which a system-level pipeline is proposed to maximize the throughput. A 720P@30FPS demo and a cross-platform demo are provided in the websites.12

Multi-view videos are becoming widely used in different fields, but their high resolution and multi-camera shooting raise significant challenges for storage and transmission. In this paper, we propose MV-MGINR, a multi-grid implicit neural representation for multi-view videos. It combines a time-indexed grid, a view-indexed grid and an integrated time and view grid. The first two grids capture common representative contents across each view and time axis respectively, and the latter one captures local details under specific view and time. Then, a synthesis net is used to upsample the multi-grid latents and generate reconstructed frames. Additionally, a motion-aware loss is introduced to enhance the reconstruction quality of moving regions. The proposed framework effectively integrates the common and local features of multi-view videos, ultimately achieving high-quality reconstruction. Compared with MPEG immersive video test model TMIV, MV-MGINR achieves bitrate savings of 72.3

In recent years, deep learning-based image compression techniques have advanced rapidly, surpassing traditional methods in terms of rate-distortion performance. However, in machine-oriented image compression, preserving high-level semantic information is of greater importance. Most existing methods employ only image-level prompts to guide frequency domain processing, leading to suboptimal preservation of semantic information for downstream machine vision tasks. To address this limitation, we propose a CLIP-guided semantic frequency domain adaptation module that extracts frequency features by applying both the fast Fourier transform and the wavelet transform. Guided by text-based semantics, the module further enhances the frequency components relevant to the target task, thereby improving machine perception performance. The proposed adapter is designed to be plug-and-play with existing learned image compression (LIC) models without requiring retraining of the full model. Experimental results demonstrate that our method outperforms state-of-the-art approaches in multiple machine vision tasks.