Search Results (853)

The evolution from 4G Long-Term Evolution (LTE) to 5G New Radio (NR) has increased cell density in cellular deployments, thereby increasing inter-cell interference (ICI), which is most significant in cell-edge scenarios. The Minimum Mean Square Error–Interference Rejection Combining (MMSE-IRC) receiver was widely adopted in LTE systems around Release 11 as a practical implementation for advanced interference rejection. Although not standardized in 3GPP specifications, it was commonly used in 3GPP studies and has been effectively used in 5G-NR environments. In this paper, we propose new versions of the Alternating Projections Hard Successive Interference Cancellation (AP-HSIC) equalizer algorithm that combine the Minimum Mean Square Error (MMSE) and MMSE-IRC equalizers (denoted MMSE-AP-HSIC and MMSE-IRC-AP-HSIC, respectively). The effectiveness of the new approach is demonstrated by Block Error Rate (BLER) analysis for Quadrature Phase Shift Keying (QPSK), 16-QAM, and 64-QAM modulations. Physical Uplink Shared Channel (PUSCH) 5G-NR receiver simulations demonstrate the differences between the proposed algorithms and the classical AP-HSIC, MMSE, and MMSE-IRC equalizer algorithms. Simulation results show that the proposed scheme, MMSE-IRC-AP-HSIC, outperforms the conventional MMSE-IRC scheme in cell-edge scenarios. Full article

Road infrastructure risk management often relies on aggregate rankings that assume stable priorities across projects. This retrospective, document-based study examines between-project differences in risk structure and directed associations between primary and secondary risks within adverse events. The analysis uses 1177 consolidated analytical records of documented cost- and/or time-related adverse events from 28 completed road infrastructure projects implemented during the construction of Pan-European Corridor X and adjacent road infrastructure in Serbia. Primary-risk, secondary-risk, and joint primary–secondary edge profiles were compared using Jensen–Shannon divergence, permutation testing, empirical-Bayes adjustment, and a beta-binomial model. Directed network analysis identified source, receiver, and bridge roles among seven risk groups. Significant heterogeneity was found in all three profile types. At the record level, 69.8% of the 1053 records containing at least one secondary risk included a secondary risk from a different group, whereas at the link level, 53.8% of the 1717 directed group-level link occurrences were cross-group. Project documentation was the dominant source and bridge, generating 59.6% of cross-group links, whereas external factors were the largest receiver and had the highest authority score. The most frequent link, from project documentation to external factors, was not overrepresented after project-specific composition was preserved. Several less frequent links were significantly overrepresented, especially those connecting project-wide factors with employer-related risks. The findings support coordinated preventive planning. The proposed data-driven framework can support the development of digital decision-support systems for project-specific risk monitoring, coordinated preventive planning, and sustainable management of road infrastructure delivery. Full article

Low-altitude UAV aerial imagery often has complex backgrounds with densely distributed small objects, posing challenges to accurate small-object detection. To address these problems, we propose a spatial–frequency response aware synergistic network for small-object detection in low-altitude UAV aerial imagery. A Frequency-Response-Aware Enhancement Module (FRAEM) is designed to effectively extract discriminative features. The module employs a deterministic stage-aware filtering strategy: Scharr-based edge-sensitive filtering is used in the shallow stage, whereas Gaussian smoothing is used in deeper stages, enabling complementary enhancement of hierarchical representations. A Detail Feature Fusion module (DFFusion) is then developed to improve the efficiency of multi-scale feature fusion. The existing Content-Aware Reassembly of Features (CARAFE) operator is employed for content-aware upsampling and feature alignment, after which DFFusion uses learnable scalar weighting to integrate high-resolution detail information with low-resolution contextual information. A Lightweight Adaptive Decoupled Head (LADH) is also designed to reduce complexity. LADH asymmetrically allocates computational capacity across the prediction tasks: the confidence branch retains stronger spatial processing, whereas the classification and regression branches use lightweight projections; depthwise separable convolution serves as an efficiency-oriented implementation choice. Experiments on the VisDrone2019 and DOTA-v2.0 datasets demonstrate that the proposed method can achieve balance between detection performance and model complexity over mainstream detection methods. Ablation experiments also prove the effectiveness of the proposed components. Full article

Holographic stereograms record large sets of multi-view projection images into elementary hologram units (hogels), and their visual quality is governed primarily by the quality and inter-view consistency of the input multi-view imagery. Conventional photogrammetry-based pipelines for generating such imagery rely on surface meshes and require burdensome post-processing, which limits the fidelity of thin structures and view-dependent reflections. This paper proposes an integrated workflow that replaces the photogrammetric stage with Gaussian splatting and presents a controlled comparison of volumetric 3D Gaussian splatting (3DGS) and surface-aligned 2D Gaussian splatting (2DGS) for hogel-based digital hologram production. Using 397 drone-captured images of a 3 m bronze Pegasus statue, both representations were trained under identical input data, camera poses, and hyperparameters, and 119,808 full-parallax multi-view images (768 × 156 grid; 120° × 52° field of view) were rendered from each model, converted into hogel arrays by an identical ray-tracing transform, and printed onto Ultimate U04 silver halide plates under identical optical conditions. In the digital domain, 3DGS outperformed 2DGS on all three standard novel-view-synthesis metrics (PSNR 32.68 vs. 31.03 dB; SSIM 0.9281 vs. 0.9139; and LPIPS 0.1150 vs. 0.1406) while using approximately 1.67 times more Gaussians. Conversely, 2DGS produced superior results at viewpoints outside the training distribution, on thin structures such as wings and mane, and at the extremes (±60°) of the virtual camera array, and these differences propagated consistently to the printed holograms, whose edge sharpness was higher for 2DGS at every measured viewpoint (mean 39.6 vs. 23.8). Notably, 2DGS achieved this superior printed-output quality while using approximately 40% fewer Gaussians than 3DGS, combining model efficiency with the multi-view consistency that proved decisive for hogel-based printing. The results demonstrate that single-view metrics such as PSNR do not capture the multi-view consistency that dominates hogel-based output quality and provide practical guidance for representation selection in holographic-printing workflows. Full article

Background/Objectives: Transthoracic lateral shoulder radiography serves as an important method for traumatic shoulder injuries and dislocations, but tissue overlap and breathing motion often compromise image quality and are particularly prominent in obese patients. This study aims to compare the image quality differences between the deep-inspiration breath-hold (DIBH) strategy and the routine free-breathing (FB) method in transthoracic lateral shoulder radiography for patients with different body mass indices (BMI), and to clarify the clinical application value of the breath-holding strategy. Methods: A total of 128 patients with traumatic shoulder injuries, glenohumeral dislocation, postoperative follow-up, and symptomatic degenerative shoulder osseous lesions were prospectively enrolled to undergo radiography. Anteroposterior (AP) projection and transthoracic lateral views applying both FB and DIBH methods were performed. They were stratified into three groups (BMI < 24 kg/m², 24–28 kg/m², and ≥28 kg/m²). Two experienced radiologists independently evaluated the subjective image quality regarding anatomical depiction and abnormality visualization between the two methods, which were compared using a four-point scale (1, poor; 4, excellent), with inter-observer consistency analyzed. The edge sharpness (ES) was also calculated and compared quantitatively. The interaction effect of BMI grouping and respiratory mode regarding ES measurement was evaluated; cumulative-link mixed-effects models with subject-level random intercepts were also fitted to test whether the benefit of DIBH varied with BMI grouping for the subjective scores. Results: The DIBH method presented superior anatomical depiction scores relative to the FB method (p < 0.001). For reader 1, median scores were 3.0 (interquartile range, IQR, 3.0–3.0) for FB radiographs and 3.5 (IQR, 3.0–4.0) for the DIBH method (z = −8.014, p < 0.001). Consistently higher median scores were also documented by reader 2 for the DIBH method: 3.0 (IQR, 3.0–3.0) for the FB method versus 3.0 (IQR, 3.0–4.0) for the DIBH method (z = −7.616, p < 0.001). For lesion depiction, the scores of the DIBH images were better than those of the FB images with p < 0.05 for reader 1 (p = 0.005) and reader 2 (p = 0.046). Moderate-to-excellent inter-reader consistency was observed. The mean ES of the DIBH method was significantly higher than that of the FB method (2.27 ± 0.73 vs. 2.02 ± 0.70; paired t = 4.80, df = 127, p < 0.001). No significant differences in ES values or subjective scores for DIBH images existed among three BMI subgroups (p > 0.05); FB anatomical scores differed significantly for both readers (reader 1: p = 0.041; reader 2: p = 0.026). There was no significant interaction between the BMI grouping and the respiratory mode regarding ES measurement (F = 0.061, df = 2, 125, p = 0.941). Conclusions: The deep-inspiration breath-hold strategy significantly improves image quality and lesion conspicuity for depicting shoulder abnormalities compared with the routine free-breathing method in transthoracic lateral shoulder radiography, without requiring additional equipment or changes to the radiographic system, when interpreted in conjunction with the corresponding AP radiograph in clinical practice for cooperative patients. Full article

Cross-organization cyber defense must reconcile collaborative learning with privacy and adversarial robustness, yet standard federated learning ships full gradient tensors, leaking sensitive posture and inviting Byzantine manipulation. We present FedMARL-LTI, a federated multi-agent reinforcement learning framework whose architecture answers both pressures with a single decision: each organization’s threat intelligence is shared only as a differentially private 768-dimensional semantic embedding, never as raw data. In the evaluated system, a Weight-DP-protected model-weight delta is also exchanged through the federated aggregator (the semantic abstraction embedding is a parallel channel); the privacy guarantee below is stated for the semantic abstraction channel, and an embeddings-only architecture—which the guarantee enables—is the design this points toward. The contribution is fourfold. (1) Semantic Abstraction (SA) channel: per organization, each round, the local gradient is summarized by an LLM, projected to a 768-dim embedding, L2-clipped, and Gaussian-noised before any numeric quantity leaves the host. The bottleneck reduces the aggregate noise magnitude—the expected L2 norm of the DP noise vector—from $O\left(\sqrt{d_{\mathrm{model}}}\right)$ to $O\left(\sqrt{m}\right)$ with $m=768\ll d_{\mathrm{model}}\approx 3\times {10}^5$. (2) Formal privacy analysis: the SA + DP cascade satisfies $(\epsilon ,\delta )$-DP and bounds per-round mutual information leakage by $\min \{T_{\mathrm{tok}}{\log }_{2}V,m/2{\log }_2(1+C^2/\left(m{\sigma }^2\right)\left)\right\}$, with Rényi composition over $T$ federation rounds. Scope of the guarantee: this bound certifies (i) the semantic-abstraction channel. It does not by itself cover (ii) the weight-aggregation channel, whose Weight-DP protection is analyzed separately, nor (iii) the whole deployed system, which is the composition of the two. We therefore state the ≈1.4-bit/MI bound as a per-round guarantee on information leaving the organization through the SA channel not over every byte the system emits; an embeddings-only configuration—which this bound enables—closes the gap to a whole-system guarantee. (3) Byzantine-resilient ClippedClustering aggregator combining L2 clipping with cosine-similarity clustering. (4) Hierarchical MARL policy with threat-profile-aware LLM-IRR reward shaping, wired end-to-end and disclosed honestly (the evaluated system uses a deterministic Johnson–Lindenstrauss projection in place of the LLM call for reproducibility; the architecture is thus LLM-compatible rather than dependent on a specific model, and a full LLM deployment is the planned extension). We evaluate on CybORG CAGE-4 with $n=5$ organizations, 30 federation rounds × 5 episodes × 100 steps per round. Releasing the SA channel in parallel shows no statistically detectable reward cost at N = 5 vs. the no-privacy baseline; this is measured at reward-shaping coefficient β = 0, so it establishes that the private semantic release does not disturb weight-channel training rather than that semantic sharing improves defense: SA-only $\Delta$reward = $+4.58$ ($t=+1.37$, NS), dual SA + Weight-DP $\Delta$reward = $+4.31$ ($t=+1.30$, NS), all $N=5$ seeds, all $\left|t\right|<1.4$. A controlled signal/noise probe confirms a $19.58\times$ improvement of SA over Weight-DP at a fixed DP budget—matching the predicted $\sqrt{d/m}\approx 19.8$. Under Byzantine sign_flip at 30% ($N=15$), ClippedClustering is directionally strongest ($F_1=0.025$ vs. FedAvg $0.020$, Krum $0.016$) but the edge is not statistically significant (CC vs. Krum $t=+1.59$, $p=0.15$, $d=+0.58$; the earlier $N=5$ “$3.4\times$” gap was small-sample optimism); its Byzantine behavior is on the harsher random_noise attack. Under a corrected implementation, the undefended baselines do not diverge or collapse; the earlier reading (Krum $0.002$, ClippedClustering $0.020$) was a noise-injection artifact and is withdrawn; ClippedClustering is now directionally best on F1 but not significantly, and trails Krum on reward (superseded Cohen’s $d=+3.77$). The cooperative-PPO family (MAPPO, IPPO) outperforms value/actor-critic (QMIX, MADDPG) by $\approx 20$ reward units, $p<0.001$. All host-level $F_1$ values stay below $0.05$ at the 15K-step training horizon used here; the relative claims of the paper (no detectable privacy reward cost, ClippedClustering’s competitive (not decisive) Byzantine behavior on the harsher attacks, cooperative-PPO dominance) are unaffected by this scope. A 200K-step long-horizon replication lifts $F_1$ above the 15K plateau (to $\approx 0.044$, $N=5$)—confirming that horizon, not the privacy/Byzantine machinery, gates absolute accuracy—but a finer 60-checkpoint run shows the climb is volatile and non-monotonic and does not reach deployment-grade, an honest stability-not-compute limitation. FedMARL-LTI is therefore presented as a proof-of-concept for the relative privacy and robustness trade-offs it isolates, not as an operationally deployable cyber defense system. We release all 141 raw run JSON outputs (Phases 1–3, the L4 backend comparison, and the algorithm/aggregator baselines), the figures, and analysis scripts for replication. Full article

Rapid urbanization and extreme rainfall have increased pluvial flood pressure in dense urban areas, yet many drainage problems emerge at microscale interfaces that conventional flood monitoring does not capture well. Focusing on the built-up area of Zhengzhou, this study uses waterlogging complaints from the People’s Daily Online Leadership Message Board to identify resident-perceived flood hotspots. Government replies, field verification, and built-environment indicators are then combined to examine how these sites are described, assigned, and validated. The complaint records reveal recurrent flood pressure at interface settings, including underground-space entrances, residential-compound–road edges, road depressions, and project boundaries. Field checks confirm that several reported hotspots correspond to visible site conditions. Government replies, however, differ in how clearly they recognize local drainage settings and responsibility boundaries. Persistent mismatches are concentrated at sites shared by multiple actors, and project-edge areas where maintenance and construction responsibilities are difficult to separate. Grid-based diagnosis further shows that planning priority is identified more effectively through the overlap of built-environment exposure, observed site pressure, and governance mismatch than through exposure indicators alone. The study treats complaint–reply exchanges as participatory spatial evidence and proposes a diagnostic procedure linking hotspot interfaces, responsibility boundaries, and field verification for pluvial flood planning. Full article

Periodic fluctuations in pump head are a common unsteady phenomenon in tubular pumps; however, their underlying energy dissipation mechanism remains insufficiently understood. This study conducted transient numerical simulations on a two-blade tubular-flow pump with highly twisted blades operating at the design flow condition. Using entropy production theory, the research quantitatively examined Rotor–Stator Interaction (RSI), vortex development, and hydraulic loss characteristics. The findings reveal that turbulent entropy production is the primary contributor to total energy dissipation, while entropy generated by wall friction is minimal. Although the impeller experiences the greatest absolute energy loss, the fluctuations in entropy production within the guide vane are significantly larger. This indicates that the energy loss represented by the entropy production in the guide vane primarily drives the periodic head fluctuations of the pump. High entropy production is concentrated near both the leading and trailing edges of the guide vane, exhibiting a trough-shaped radial distribution influenced by the leading-edge hub vortex and tip leakage vortex. Additionally, the transient changes in entropy production under RSI are governed by the periodic formation and strengthening of the guide vane passage vortex, along with the shedding and breakdown of the wake vortex. Velocity analysis shows that the circumferential movement of the impeller wake continuously modifies the instantaneous inflow conditions at the guide vane’s leading edge, causing periodic changes in the incidence angle that enhance the passage vortex while weakening the wake vortex. This study provides deep insights into the operational stability of pumps and pumping stations in water transfer projects. Full article

Petroleum engineering projects face complex risk environments. Existing risk control systems often only provide overall risk safety thresholds. They lack effective quantitative interval estimation methods. Moreover, translating quantitative analysis findings into actionable on-site management instructions remains challenging. Symmetry serves as a core analytical perspective for cutting-edge research in control theory and system engineering. Feedforward and feedback controls are functionally complementary and sequentially cascaded, featuring intrinsic complementary symmetry. From the perspective of feedforward–feedback symmetry, this study constructs a GA-LLM hybrid framework combining genetic algorithm (GA) and large language models (LLMs) to address the above shortcomings. This framework is jointly composed of four collaboratively functioning modules, comprising the risk status input module, GA feedforward control module, retrieval-augmented generation (RAG) knowledge retrieval module, and the LLM feedback control strategy-generation module. In this framework, the GA serves as the feedforward controller, which computes the joint inscribed control box for each risk factor offline based on the risk relationship model established by the Back Propagation (BP) neural network. The RAG-enhanced LLM serves as the feedback controller, dynamically generating structured risk control instructions based on deviations. Case validation results demonstrate that the BP neural network achieved excellent performance with an R² of 0.99575 under leave-one-out cross-validation. The GA successfully solved the joint control box for the 14 risk factors, achieving a 100% joint constraint satisfaction rate for any combination within the box. The RAG retrieval module achieved a Recall@5 of 0.8933, MRR@5 of 0.7367, nDCG@5 of 0.7505, and Success@5 of 1.000. Ablation experiments show that the RAG-LLM scheme outperformed both the LLM without RAG scheme and the rule-based template scheme across four dimensions, with an inter-rater reliability ICC(2,1) of 0.719, reaching a moderate reliability level. This study integrates the quantitative optimization capability of GA with the semantic generation capability of LLM, enabling the transformation from the joint control box to executable management instructions. It not only provides a practical tool for petroleum engineering risk management but also offers new insights for the design of intelligent control systems from the perspective of feedforward–feedback symmetry. Full article

Rapid and unplanned urbanization in Dhaka is reshaping land use, intensifying peripheral expansion, and increasing pressure on urban and ecological resources. Understanding these growth dynamics is essential for effective urban growth boundary delineation and sustainable planning, yet integrated assessments of historical and future urban growth remain limited. The current study evaluates spatiotemporal urban expansion from 2010 to 2025, delineates the urban growth boundary using a morphological framework, and simulates future growth for 2030 through a coupled cellular automaton-neural network model. Multi-temporal Landsat imagery (2010, 2015, 2020, and 2025) was classified in Google Earth Engine using supervised Maximum Likelihood Classification. Urban growth patterns were quantified using the urban expansion intensity index (UEII), annual urban expansion rate (AUER), and landscape expansion index (LEI). The urban largest continuous patch index (ULCPI) approach was applied to extract functional urban boundaries. Model performance was validated using the Chi-square (χ²) goodness-of-fit test. Results show a substantial increase in built-up land from 115.85 km² to 171.42 km² between 2010 and 2025, accompanied by a decline of approximately 60 km² in urban green spaces. LEI results demonstrate a transition from compact infilling growth (2010–2015) to dominant edge and outlying expansion (2015–2020), indicating progressive peri-urbanization. The urban largest continuous patch (ULCP) nearly doubled from 78.58 km² to 152.13 km² over the same period, accentuating rapid spatial consolidation. The 2030 projection anticipates continued corridor-oriented expansion, particularly toward the northern and eastern peripheries, with predictive agreement from the CA–ANN model (χ² = 0.03 < 7.8). The study identifies a clear transition from monocentric compactness to polycentric expansion, emphasizing the necessity for enforceable growth containment, transit-oriented development, and ecologically responsive planning strategies to ensure long-term urban sustainability. Full article

Top-loading smart vending cabinets require precise transaction-level product detection under strict hardware and deployment constraints. In this paper, “sensorless” refers specifically to the absence of auxiliary product-level sensing hardware, such as RFID tags, weight sensors, shelf load cells, or product-slot instrumentation; the system still uses two camera sensors. Conventional snapshot-difference methods compare only a small number of frames at the beginning and end of a transaction and therefore cannot explicitly represent intermediate product motion, such as pickup, return, inspection, occlusion, and shelf resettling. This paper proposes ZAB-Fusion, a dual-camera edge sensing framework with zone-aware multi-object tracking for sensorless smart vending cabinets. The framework combines a YOLO11-seg and RT-DETR detection ensemble with ByteTrack temporal association, projects product tracks into a three-zone vertical cabinet model, interprets compressed zone sequences using a finite-state event classifier, and integrates camera-specific event streams through an evidence-gated cross-camera fusion rule. The proposed method was evaluated on 220 in-service vending transactions containing 227 ground-truth TAKEN events and 87 RETURNED events across seven product classes. Compared with the snapshot-difference baseline, ZAB-Fusion improved recall from 0.665 to 0.925 and F1-score from 0.780 to 0.944, while maintaining a high precision of 0.963. At the transaction level, exact receipt accuracy increased from 0.645 to 0.900. Runtime analysis on an Intel N100 CPU-only edge device showed an average processing latency of 562 ms per transaction under the selected-frame inference protocol. The zone classification, finite-state event interpretation, and evidence-gated cross-camera fusion stages required only 3 ms in total. These results demonstrate that explicit motion semantics and auditable cross-camera evidence gating can improve sensorless retail transaction level recognition in sensorless smart vending cabinets without adding auxiliary product-level sensing hardware. Full article

Efficient task offloading in UAV-assisted heterogeneous mobile edge computing (MEC) networks is increasingly challenged by the co-existence of operationally distinct workload scenarios—including high-demand bursts, resource-constrained periods, and balanced operational states—each demanding fundamentally different assignment strategies. In such networks, mobile executor nodes (e.g., UAVs or vehicle-mounted edge servers) must be dispatched to the vicinity of geographically distributed tasks, making assignment decisions jointly dependent on node mobility, the quality of sensing data, and dynamic resource availability. Conventional approaches based on combinatorial optimization with fixed parameters or greedy heuristics fail to adapt to these varying conditions, leading to resource depletion under sequential workloads or underutilization under high-demand bursts. To address these limitations, this paper proposes SAGE (Scene-Adaptive Graph-Enhanced offloading), a task-offloading framework that combines a heterogeneous graph neural network (HeteroGNN) with a dueling double DQN meta-controller and a mixed-integer linear programming (MILP) solver. At the state-representation level, a heterogeneous bipartite graph is constructed over mobile executor nodes and tasks, with type-specific projection layers encoding the semantic features of each node type and three-dimensional edge features—comprising task success probability, normalized service distance, and link quality—integrated via edge-gated message passing. At the decision level, the meta-controller perceives the current workload scenario through a seven-dimensional situational state vector fused with the graph embedding, selects an appropriate offloading strategy from a learned discrete action space, and drives the MILP solver to perform task-chain assignment under the selected configuration. Experiments on 60 fixed evaluation episodes spanning three representative workload scenarios demonstrate that SAGE achieves an overall reward improvement of 15.9% over the best fixed-strategy baseline, reduces the resource depletion rate to 16.7%, and maintains a high-priority task completion rate of 84.7%. Particularly under resource-constrained conditions, SAGE reduces the reward deficit by 72.3% relative to the best fixed strategy (from $-0.531$ to $-0.147$), demonstrating strong scene-adaptive decision-making capability. Full article

Green-building research has examined energy efficiency and indoor environmental quality in depth, whereas site-related strategies are often represented through aggregate certification outcomes with limited visibility into criterion-level evidence and weighting assumptions. This study develops the Climate-Referenced SS–LT Site-Performance Assessment Framework (CR-SSAF) for an exploratory documentation-based comparison of a combined Sustainable Sites, Location, and Transportation (SS–LT) construct across five green-certified smart office buildings in three Köppen climate zones. Six SS–LT criteria were assessed using a four-level operational rubric, an exact-normalized weighted Site-Performance Index (SSPI), a descriptive Technology Enablement Factor (TEF), ordinal inter-rater agreement analysis, three weighting schemes, and a TOPSIS ranking-concordance check. SSPI values ranged from 2.00 to 2.65. The Af case recorded the highest SSPI in this sample, while the lowest scores occurred where no qualifying project-specific heat-mitigation or green-/open-space evidence could be verified. The first, second, and last ranks remained unchanged across the three weighting schemes, while The Edge and Shanghai Tower were tied under the equal and ecology-sensitive schemes. Because the sample is small and heterogeneous, and does not control for urban form, building scale, infrastructure, certification system, or documentation availability, differences cannot be attributed independently to climate. CR-SSAF is therefore presented as a transparent exploratory workflow, not as a validated climate effects model or as evidence of transferability. Full article

With the accelerated progress of computational aesthetics and digital interior design, the demand for real-time and precise detection of aesthetic objects on edge devices has become increasingly pressing in applications such as intelligent design assistance, domestic aesthetic assessment, and augmented reality-based interior staging. As a core task in digital home aesthetics governance, virtual interior furnishing, household cultural archive development, and automated aesthetic evaluation, multi-view aesthetic object detection plays an essential role. However, this task still faces substantial difficulties arising from pronounced viewpoint variation, scale inconsistency, reflective materials, intricate decorative patterns, and cluttered indoor scenes. To address these issues, this study presents GhostNetV2-YOLO, a lightweight yet robust detection framework designed for accurate localization of aesthetic objects under unconstrained multi-view acquisition settings. The task is formally defined as closed-set detection of 10 pre-selected home aesthetic decorative items, including both planar decorative pieces and three-dimensional ornamental objects, and all performance claims are bounded within the horizontal bounding box detection paradigm. The framework incorporates three complementary components tailored to the target task. First, a task-adapted GhostNetV2 backbone is employed to enable efficient multi-scale feature extraction and long-range dependency modeling, with optimization specifically oriented toward structured aesthetic objects with stable global contours under viewpoint variation. Second, an improved Attention-based Intra-scale Feature Interaction (AIFI) module is introduced, integrating compressed QKV projection, linear attention, depthwise spatial refinement, and channel gating so that reflection-induced noise and background disturbance can be effectively reduced. Third, an enhanced Distance-IoU regression loss is adopted, in which explicit edge alignment and dynamic sample weighting are incorporated to improve boundary regression accuracy for rectangular and regularly contoured aesthetic objects. These designs jointly enhance contextual representation, boundary localization, and computational efficiency. Extensive experiments on two newly constructed multi-view aesthetic object datasets (AestheticHome-12K and AestheticHome-2K) demonstrate that the proposed detector achieves 94.80 ± 0.32%/94.20 ± 0.37% mAP@0.5, 96.30 ± 0.28%/95.60 ± 0.31% precision, and 94.70 ± 0.35%/93.80 ± 0.39% recall across two datasets (reported as mean ± standard deviation of 5 independent training runs with distinct random seeds), with only 2.89 M parameters and 6.0 GFLOPs. Statistical significance is verified via paired two-tailed t-tests with Bonferroni correction (adjusted p < 0.05) for all performance comparisons against baseline models. Compared with the YOLOv11n baseline, the method improves mAP@0.5 by 1.87–2.09 percentage points and recall by 3.27–3.48 percentage points while reducing computational cost. Notably, it also achieves 79.2–80.5% mAP@0.5:0.95, outperforming the baseline by 4.7–4.9 percentage points, indicating significantly superior localization accuracy under stricter criteria. The proposed model achieves a remarkable balance between accuracy and efficiency, making it highly suitable for deployment on resource-constrained edge devices commonly used in digital design and home aesthetic monitoring systems. The results indicate that combining lightweight long-range feature extraction optimized for rigid aesthetic objects, compact attention-based feature interaction for interference suppression, and geometry-aware regression tailored for aesthetic targets provides an effective and efficient solution for robust aesthetic object detection in real-world computational aesthetics and digital interior design applications. Full article

(1) Background: Reliable forehead region-of-interest (ROI) localization remains challenging in low-resolution thermal imaging because limited spatial detail increases localization uncertainty. This study develops a lightweight nostril-anchored geometric ROI projection method for contactless forehead skin temperature estimation. (2) Methods: The forehead ROI was projected from the detected nostril bounding box using anthropometry-guided proportional geometry, followed by 95th-percentile ($P_{95}$) spatial aggregation and exponential moving average (EMA) stabilization. Direct localization was evaluated against fixed-crop and MediaPipe Pose-based baselines using 120 annotated pilot frames across 0.5–1.5 m, followed by validation on 150 additional annotated frames from ten formal participants. A medical non-contact infrared forehead thermometer targeting the same central forehead region was used as a practical, rather than traceably calibrated, comparator. (3) Results: The proposed method outperformed the two localization baselines in the pilot evaluation. Formal-cohort validation achieved an overall IoU of 0.714, FIR of 0.870, and NCE of 0.091, with a 0.7% spatial-failure rate. Across 90 temperature measurements, the method achieved an MAE of $0.514 °\mathrm{C}$, RMSE of $0.614 °\mathrm{C}$, and mean bias of $-0.187 °\mathrm{C}$, with 95% Bland–Altman limits of agreement from $-1.340 °\mathrm{C}$ to $0.966 °\mathrm{C}$. Real-time implementation achieved $24.12\pm 1.43$ FPS on a Jetson Orin Nano. (4) Conclusions: The proposed method provides a computationally efficient forehead ROI localization approach for low-resolution thermal imaging. Localization was strongest at 0.5 m and lower at farther distances; therefore, its scale adaptation should not be interpreted as distance-invariant performance. The temperature results represent agreement with the practical comparator rather than absolute clinical accuracy. Full article