Dong Wang

Artifacts, including halos and saturated bright spots, are common in spaceborne optical images and can degrade the reliability of star extraction, space object detection, and photometric analysis. Traditional signal-processing methods, such as morphological filtering and low-rank decomposition, rely on fixed priors that may fail under complex artifact morphologies. Deep restoration networks can improve visual quality but do not explicitly enforce radiometric consistency. Multimodal instruction-driven editing models provide semantic localization capability, but probabilistic diffusion resampling can introduce uncontrolled pixel changes in non-target regions, compromising pixel-level physical consistency. We refer to this problem as editing-induced radiometric drift. To address this problem, we propose PARE, a physics-preserving attention-guided artifact removal framework for spaceborne optical images. Instead of directly using the edited image as the final restoration, PARE treats it as a candidate restoration and derives artifact-region constraints from the image-to-text cross-attention sub-block of the MM-DiT joint attention matrix. These constraints are combined with multi-scale fusion to restrict generative modification to localized artifact regions, thereby enabling artifact suppression while reducing unintended changes in non-target regions. Experiments on simulated and real spaceborne optical image datasets show that PARE achieves effective artifact suppression while improving radiometric preservation. On real on-orbit images, PARE reaches 92.95% artifact mean reduction (AMR) and 98.57% artifact energy reduction (AER), reduces the outer-region mean absolute error (O-MAE) to 0.54, and improves the outer-region structural similarity index (O-SSIM) to 0.997. It also yields the smallest background shifts among all compared methods. These results indicate that PARE provides a favorable trade-off between artifact suppression and radiometric fidelity and offers a practical way to apply generative models to scientific imaging tasks that require pixel-level physical consistency.

With the rapid advancement of Transformer-based large language models (LLMs), these models have found widespread applications in industrial domains such as code generation and non-functional requirement (NFR) classification in software engineering. However, recent research has primarily focused on optimizing linear matrix operations, while nonlinear operators remain relatively underexplored. This paper proposes hardware-efficient approximation and acceleration methods for the Softmax and RMSNorm operators to reduce resource cost and accelerate Transformer inference while maintaining model accuracy. For the Softmax operator, an additional range reduction based on the SafeSoftmax technique enables the adoption of a bipartite lookup table (LUT) approximation and acceleration. The bit-width configuration is optimized through Pareto frontier analysis to balance precision and hardware cost, and an error compensation mechanism is further applied to preserve numerical accuracy. The division is reformulated as a logarithmic subtraction implemented with a small LOD-driven lookup table, eliminating expensive dividers. For RMSNorm, LOD is further leveraged to decompose the reciprocal square root into mantissa and exponent parts, enabling parallel table lookup and a single multiplication. Based on these optimizations, an FPGA-based pipelined accelerator is implemented, achieving low operator-level latency and power consumption with significantly reduced hardware resource usage while preserving model accuracy.

Recently, post-training quantization (PTQ) methods for large language models (LLMs) primarily focus on tackling the challenges caused by outliers. Scaling transformation has proven to be effective while how to enhance the performance of extremely low-bitwidth (e.g., 2-bit) PTQ under it remains largely unexplored. In this work, a new PTQ framework, namely MPPQ, is established. Specifically, MPPQ first proposes an enhanced reconstruction loss based on Mixed metric supervision to mitigate the distribution inconsistency caused by quantization while providing strong regularization for learnable parameters. Secondly, we introduce a Proxy-based adaptive rounding scheme in weight quantization, which replaces the round-to-nearest (RTN) function to minimize the overall quantization errors through element-wise scaling. Furthermore, a factor coarse Pre-searching mechanism is presented to ensure proper coordination between quantization and clipping patterns, while achieving optimal initialization of clipping factors before training. Extensive experiments show that MPPQ consistently outperforms state-of-the-art methods in low-bit quantization settings. For instance, the perplexity of WikiText2 can be dramatically reduced to 8.85 (3.9 ↓ vs 12.75 of the latest method, LRQuant) for the LLaMA-2-7B model, which is quantized with W4A4.