The Transformer-based DETR framework has achieved end-to-end set prediction, driving technological innovation in the field of object detection. However, it has two key limitations: First, the cost function overlooks semantic similarity between queries and targets, and the Hungarian algorithm fails to capture their intrinsic relationships, resulting in inaccurate matches. Second, the training phase imposes positional constraints only on positive samples and lacks an exclusion mechanism for neighboring samples, limiting detection performance. This paper analyzes the spatial proximity and semantic similarity between neighboring negative samples and positive samples, which weakens the model's ability to distinguish between positive and negative samples. To address this, a Neighboring-Aware Optimization Strategy (NAOS) is proposed and applied to the DETR framework, resulting in the NAO-DETR method. The proposed method enhances the Hungarian matching process by incorporating a semantic similarity cost between queries and targets, thereby alleviating query–target matching ambiguity. In addition, it integrates explicit supervision through geometric and coverage constraints on neighboring samples, which effectively suppresses positional interference during regression and consequently improves the accuracy of object localization. Experimental results demonstrate that NAO-DETR achieves superior performance, surpassing baseline models DINO and Deformable-DETR by 2.1% and 5.7% in mAP on COCO datasets, respectively. Moreover, NAOS demonstrates strong generalization, effectively transferring to DETR framework to improve its detection performance.
Robust Reversible Data Hiding (RRDH) predominantly focuses on compression, with insufficient attention to geometric attacks. Existing geometric attack-oriented schemes are confined to 1-bit watermark capacity, failing to meet practical demands. To address this, this paper proposes a novel wide-domain RRDH scheme based on synergy between singular value differencing in multi-directional block-pairs and pre-extracted residuals. First, a mechanism constructs adjacent block-pairs across multiple directions and performs singular value differencing. It extracts maximum singular values via Singular Value Decomposition (SVD) and embeds 4-bit watermarks via differential modulation, and resists ±30° rotation, 0.8×–1.5× scaling, and 2-pixel translation. Second, a texture-adaptive modulation strategy dynamically adjusts singular value modification amplitude with image gradient magnitude, maintaining robustness while reducing host image distortion. After embedding, the average host image PSNR remains ≥35 dB. Third, a residual-watermark synergistic embedding framework pre-extracts residual matrices, compresses them, and adaptively embeds them via least significant bit (LSB) modification within reasonable capacity limits, achieving complete lossless recovery of the host image in attack-free scenarios. Experiments verify excellent watermark extraction accuracy under these geometric attacks and Joint Photographic Experts Group (JPEG) compression (quality factor ≥80), effectively resolving key limitations in existing schemes, including multi-directional geometric attack adaptation, robustness-quality balance, and embedding-extraction compatibility.
The ever-growing number of energy requirements needed in cloud data centers has led to the rising use of intelligent, sustainable and environmental scheduling algorithms. The study proposes a new hybrid algorithm that is a fusion of Proximal Policy Optimization (PPO) and Round Robin (RR), and is called PPO-RR, to enhance the efficiency of virtual machine (VM) scheduling and reduce the amount of energy consumption in a virtualized cloud system. The suggested approach follows the strategy of reinforcement learning: using the feedback about the system state, it can actively optimize the scheduling decisions to improve resource utilization and eliminate unnecessary waste of power in idle conditions.Experiments were carried using an artificially created cloud environment dataset. The PPO-RR model was seen to have shown significant performance gains over the conventional RR and heuristic scheduling techniques. To be more precise, PPO-RR achieved the accuracy of 95.3%, precision of 94.6%, recall of 95.4, and F1-score of 94.9. Such findings demonstrate the superiority of the model in its capacity to assign the VMs to the hosts appropriately during different workloads.With regard to energy measures, PPO-RR saved 1520 kWh to 1235 kWh translating to about 56.3 percent of energy saved as compared to conventions practices. The PPO-RR confusion matrix shows that the True Positives (TP) are 420, True Negatives (TN) are 360, False Positives (FP) are 15 and False Negatives (FN) are 25 which means that the accuracy of the classification of the decisions in scheduling is high.The model offers dynamic scheduling with performance and energy benefits compared to the traditional schedulers because the PPO-RR model adapts to the work load profile and system conditions in a real-time fashion. Despite the fact that this study shows PPO-RR to positively perform in various metrics, no formal statistical test of significance (e.g., t-test) was conducted and will be done in subsequent research.In sum, the PPO-RR approach provides a flexible and extensible mechanism to implement sustainable cloud computing, since it contributes to satisfying the objectives of green computing with the help of well-organized schedules.
Background Intangible Cultural Heritage (ICH) Chinese patterns encode millennia of cultural wisdom through complex, multi-layered semantic structures as well. Current computational approaches are unable to break down or recreate these patterns with cultural authenticity. Objective This paper presents MSD-CLIP (Multi-Granular Semantic Deconstruction and Morphological Regeneration CLIP), a new framework that aims at systematically decomposing Chinese ICH patterns into interpretable cultural values and reassembling them with accurate semantic control, without losing traditional integrity. Methods MSD-CLIP combines a hierarchical semantic deconstruction framework that examines patterns at five granularity levels: Dynasty-Temporal, Wu Xing Philosophy, Regional-Cultural, Symbolic-Functional, and Aesthetic-Compositional. It has a controllable semantic regeneration engine, which allows fine-grained manipulation of single cultural attributes, and a cross-granular consistency enforcer that ensures that regenerated patterns conform to the traditional criteria. Results Five-fold cross-validation validated model robustness (SD < 0.5%, p < 0.001 vs baselines). MSD-CLIP was tested on 3247 expert-annotated Chinese ICH patterns and demonstrated superiority over existing models on all measures. It attained 91.3 +/- 0.36% deconstruction precision, 88.7 +/- 0.42% regeneration fidelity, 4.2 +/- 0.10 cultural authenticity score, and 87.4% control accuracy. The accuracy of multi-granular deconstruction was high across all semantic levels, ranging from 87.9%-91.3%. Controlled regeneration achieved an average accuracy of 87.5% in attribute manipulation. Professional assessment indicated 84% overall approval, reliably preserving cultural authenticity. Conclusion MSD-CLIP is the first computational instrument that can deconstruct and regenerate the semantic Chinese ICH patterns at a systematic level with morphological re-inspiration. It offers a major contribution in digital heritage preservation, culturally-conscious design generation, and educational uses, without compromising the cultural integrity.
