Computer Science and Mathematics

Abstract: Multi-UAV emergency monitoring can enlarge spatial coverage, but overlapping fields of view create dependent detections that violate the conditional-independence assumption of conventional Bayesian accumulation. This study presents a correlation-aware method that converts georeferenced neural-detector outputs from multiple unmanned aerial vehicles (UAVs) into a calibrated probabilistic emergency map. Detector likelihoods are empirically calibrated, projected detections are distributed over a spatial grid, and repeated observations are down-weighted through an effective sample size derived from viewpoint and temporal similarity. The method is evaluated with a geometrically exact camera simulation and a detector emulator calibrated from 10,463 real annotated frames. In a synthetic model with an exactly computable posterior, correlation-aware fusion reduces the median excess Brier score from 23.4% for naive fusion to 3.9%, and the worst case from 91.5% to 8.2%. In the flight simulation, expected calibration error decreases from 0.046 to 0.013 for tightly clustered viewpoints. A weighted Voronoi coordination layer additionally reduces coverage time by a factor of 3.47 and duplicated imaging from 100% to 10.0% over 1 km² with five UAVs. The results show that explicitly accounting for dependence is important for probabilistic multi-UAV situational awareness.

Abstract: This paper presents a prototype of a smart security system for residential complexes that integrates real-time facial recognition, aggressive body language analysis, and weapon detection. Its service architecture utilizes a backend featuring hybrid persistence (PostgreSQL, pgvector, and MinIO) alongside WebSockets for immediate alert dissemination. A centralized rules engine correlates detections from individual modules (faces, poses, and YOLO models) into actionable incidents, thereby optimizing response capability and residential surveillance.

Abstract: Agentic artificial intelligence marks a shift from static prediction toward systems that perceive, reason, act, verify, and adapt through closed-loop interaction with external environments. Although large multimodal models provide the representational substrate for this shift, practical agency emerges only when model inference is embedded within memory, planning, tool use, environment interfaces, verification, and oversight mechanisms. We formalize the notion of an agentic AI system, distinguish it from single-turn assistants, scripted workflows, retrieval-augmented generation, and classical embodied controllers, and propose a taxonomy covering model substrates, observation spaces, memory, planning, action interfaces, multi-agent coordination, evaluation, and autonomy constraints. We then synthesize learning and optimization paradigms for agents, including supervised trajectory tuning, tool-use adaptation, preference optimization, reinforcement learning, process rewards, verifier-guided search, test-time scaling, memory-driven adaptation, and distillation. Across domains such as web search, software engineering, multimodal reasoning, embodied control, scientific discovery, and enterprise automation, we analyze the factors shaping agent performance and reliability. We further examine benchmarks, evaluation protocols, safety considerations, governance challenges, and open research problems in long-horizon reasoning, memory management, multimodal grounding, and trustworthy deployment. A companion repository tracking developments in agentic AI is available at https://github.com/Tajamul21/awesome-agentic-ai.

Abstract: Automated detection of visual indicator targets associated with selected phenological stages can support image-based monitoring of greenhouse cucumber. However, female flowers and young fruits are often small, partially occluded, and visually similar to adjacent developmental indicators, which complicates their detection in greenhouse scenes. In this study, YOLO11n-Cucumber was developed from YOLO11n for detecting seedling plants, female flowers, young fruits, and marketable fruits. The model incorporates an MDGC-GhostNetV3 backbone for feature extraction, a C2PSA-SCSA module for feature enhancement, and an MSF-CIoU loss for bounding-box regression. Experiments were conducted on a greenhouse cucumber dataset constructed in this study, comprising 3,849 images and 8,791 annotated instances. On the internal test set, YOLO11n-Cucumber achieved a precision of 86.4%, a recall of 84.2%, an mAP@0.5 of 86.8%, and an mAP@0.5:0.95 of 61.2%. Compared with YOLO11n, mAP@0.5 and mAP@0.5:0.95 increased by 2.3 and 4.1 percentage points, respectively, while floating-point operations (FLOPs) decreased from 6.4 G to 5.0 G. This reduction in FLOPs was accompanied by increases in the parameter count, from 2.6 M to 3.2 M, and in model size, from 5.6 MB to 6.8 MB. These results indicate that YOLO11n-Cucumber provides an improved accuracy–FLOPs trade-off under the greenhouse image-acquisition conditions of this study. Its generalizability across cultivars, greenhouse types, devices, and environmental conditions requires further evaluation.

Abstract: Image fusion combines images from different sources into a single, more informative image and is used in panorama construction, remote sensing, medical imaging and infrared–visible surveillance, among other fields. The quality of the fused result depends directly on how accurately the input images have been geometrically aligned (registered) beforehand. This paper systematically presents two complementary approaches for aligning two images for fusion: (i) feature-based alignment for images from similar sensors (SIFT/SURF/ORB, ratio test, RANSAC, homography), and (ii) intensity-based alignment using mutual information (MI) for images acquired by different sensors at different resolutions. Preprocessing, transformation model selection, optimization, warping, fusion and post processing are organized into an end-to-end pipeline, and practical guidance is given on handling resolution differences in physical coordinates. The paper also discusses frequent implementation pitfalls in MI-based registration and appropriate evaluation measures. The pipeline is described reproducibly using the open-source OpenCV and SimpleITK libraries.

Abstract: Many food products carry a decoration that is placed at random and is always normal, such as sesame seeds on a burger bun, while the defects that matter lie on the surface underneath. We call this setting stochastic normal decoration and introduce a benchmark for it: plain loaves with and without synthetic sesame placed by a model fitted to real buns, so that every defective loaf exists in a paired form, together with real sesame buns. On it, PatchCore loses most of its detectability and its heat concentrates on the seeds. A 2x2 swap of memory bank and test image shows why: at the defect pixels that plain PatchCore finds, sesame leaves the nearest-neighbour distance unchanged; it raises the distance of normal surface, most strongly next to seeds. We therefore propose decoration-context normalisation, which standardises each pixel's PatchCore distance against normal pixels that share its context (distance to the nearest detected seed, local seed density, distance to the region edge), using normal calibration images outside the memory bank. It needs no defect labels or retraining and removes no pixel. With calibration units physically disjoint from bank and test, it raises image AUROC on the paired set from 0.768 to 0.920 (all six loaves improve) and on real buns from 0.894 to 0.945 (four of five buns; not resolved with five buns). Score-level masking of detected seeds reaches numerically similar AUROC but stops inspecting up to 30% of the bun and, in a stress test, detects white seed-shaped foreign matter less well; both fail on piles of seeds. In a synthetic black-sesame transfer test the method again improves every loaf over PatchCore, while masking is better when the decoration detector also picks up dark defects.

Abstract: Large collections of Uzbek books and dissertations exist only as scans, and general-purpose OCR confuses the Uzbek Cyrillic letters ў, қ, ғ and ҳ with their Russian look-alikes. We present uzocr-8b, an open full-page OCR model for Uzbek Cyrillic and Latin obtained by staged LoRA adaptation of Qwen3-VL-8B-Instruct on exactly labelled synthetic pages and pseudo-labelled real scans. We formalise the model and training objective, a repetition-loop criterion and retry policy, and a conservative lexicon rule for the four Uzbek-specific letters. We describe a page-level benchmark of 300 pages (61,370 reference words) from 11 held-out typeset and letterpress Uzbek documents printed between 1999 and 2023, with a reference checked against the scans, and an evaluation protocol with fixed normalisation, corpus-level error rates and bootstrap confidence intervals. On the 284 pages retained by a coverage filter, uzocr-8b reaches a normalised word error rate of 1.42% without post-processing, 1.74% on Cyrillic and 0.65% on Latin pages, against 11.98% for its base model, 7.91% for Qwen3.6-35B-A3B and 25.64% for Tesseract 5; on all 300 pages, 4.03% against 14.65%, 10.16% and 27.53%. The synthetic-only checkpoint reaches 3.17%. One RTX 5090 processes a page in about 1.2–2.0 s. The weights are released under Apache-2.0.

Abstract: AI-generated video often fails in time rather than in any single frame. Typical failures are global brightness flicker, frozen segments, padded duplicate frames, high-frequency texture “boil”, and abrupt discontinuities. Learned quality models can score such clips, but they need GPU inference, pinned checkpoints, and feature extractors that change between releases, which makes results hard to reproduce outside the original lab. This paper proposes TSM, a training-free Temporal Stability Metric suite that needs only FFmpeg and array arithmetic. TSM computes six interpretable measures on downsampled luma: activity, freeze ratio, flicker index, temporal jitter ratio, a flash-robust shot-boundary count, and a native-rate duplicate-frame ratio. It also defines a frame-difference time series that serves as a temporal signature for detecting re-encoded or re-framed variants of the same clip. We validate TSM in three ways. First, controlled perturbations of 15 clean clips show that each target metric responded in the expected direction in 15 of 15 clips for freezes, flicker, frame duplication, and temporal noise, and in 14 of 15 for an inserted hard cut. Second, on 57 candidate shot boundaries annotated by visual inspection, the TSM detector reached precision 0.756 and recall 0.944, against 0.875 and 0.972 for FFmpeg's scdet; requiring both detectors gave 0.917 for both. Third, re-encoding at CRF 35 changed every continuous metric by a median of at most 3.4%. We then apply TSM to 35 public showcase files from three AI video websites (Section 6). Checksums show that 10 files are byte-identical across two sites, which leaves 25 unique clips, and the temporal signature identifies one portrait/widescreen variant pair. The case study shows how TSM separates multi-shot edits from single-shot clips before any quality comparison. It does not rank the sites or the generators behind them.

Abstract: Accurate and timely identification of cauliflower diseases is essential for reducing crop losses and supporting effective disease management. Although deep learning has shown strong performance in plant disease classification, comprehensive comparisons of convolutional neural network (CNN) and Transformer architectures across different image resolutions and augmentation settings remain limited for cauliflower disease recognition. This study presents a benchmark evaluation using a harmonized dataset of 5,265 original cauliflower images distributed across 15 classes. Following stratified partitioning, offline augmentation increased the training subset from 3,703 to 7,980 images, while the validation and test subsets each contained 781 images. Six CNN architectures (ResNet50, DenseNet121, MobileNetV3-Large, EfficientNetV2-S, Xception, and ConvNeXt-Tiny) and two Transformer-based models (ViT-Base-Patch16 and Twins-SVT-Base) were fine-tuned using ImageNet-pretrained weights. Six experimental configurations combined input resolutions of 224 × 224, 300 × 300, and 384 × 384 with and without additional online augmentation. Performance was evaluated using accuracy, precision, recall, F1-score, confusion matrices, and Grad-CAM visualizations. Twins-SVT-Base achieved the highest observed test accuracy under the 384 × 384 configuration with additional online augmentation, reaching 98.21% accuracy and a weighted F1-score of 98.17%. ViT-Base-Patch16-384 achieved its highest performance without additional online augmentation, obtaining 97.82% accuracy and a macro F1-score of 97.95%. Performance varied across image resolutions, augmentation settings, and model architectures. Additional online augmentation was associated with the largest observed improvement for Twins-SVT-Base, whereas ViT-Base-Patch16-384 performed best without it. Overall, the findings highlight the potential of modern CNN and Transformer architectures for cauliflower disease classification and emphasize the importance of architecture-specific optimization of image resolution and augmentation strategies.

Abstract: Many industrial vision questions are binary ("is this suitcase damaged?") and must be answered on premises, with a confidence that can be trusted and an option to defer to a human. We study how far small open vision-language models (VLMs) get on such questions and what it takes to make them reliable. First, a pre-registered benchmark of 12 open VLMs (0.8B-12B parameters; 1,807 yes/no items in four question families; test split analysed once) on an NVIDIA Jetson AGX Orin finds flat discrimination among the Qwen-based models of 2B parameters and above (pooled test AUROC 0.930-0.940), no gain from models post-trained for decisions over their own base models, and calibration that varies far more across models than ranking does: raw expected calibration error ranges from 0.014 to 0.122, and one shared Platt step brings every model to 0.013-0.037. Qwen3-VL-2B is the fastest model (0.45 s per image, 4.4 GB), and no model was significantly better than it in any tested comparison. Second, we attach per-task heads, L2-regularised logistic regressions on the frozen model's last-layer hidden state, and abstain below a confidence threshold set by the application. On test photos from sources unseen in training, a suitcase-damage head raises AUROC from 0.925 (calibrated model answer) to 0.962, paired 95% CI of the gain [+0.018, +0.058]; at a 0.90 threshold it answers 71.7% of the photos with 96.6% accuracy, against 61.1% for the calibrated answer. A baggage-presence head does not improve AUROC. Third, a case study runs the calibrated model on an 8 GB laptop: on 148 test photos it decides 49% automatically at 97.3% accuracy and leaves the rest to staff, at no API cost. Limitations: internet photos only, four question families, and a source confound in one damage test set.

Abstract: Phylogenetic trees are routinely inspected, rooted, annotated and illustrated in comparative genomics, molecular epidemiology and microbial ecology, yet existing tools force a trade-off between interactive desktop editing, layered metadata annotation, data privacy and reproducible figure generation. Here we present PhyloGPU, an open-source desktop application for viewing, editing and annotating phylogenetic trees. PhyloGPU reads Newick, NEXUS and PhyloXML files and displays them in rectangular, slanted, circular and unrooted layouts with continuous zoom. Trees can be edited interactively (re-rooting, midpoint and outgroup rooting, rotation, ladderizing, collapsing, pruning and renaming) and can be built from alignments or distance matrices by neighbor-joining or UPGMA. Tabular metadata are drawn as iTOL-style tracks (colour strips, heatmaps, bar charts and text) around circular trees or beside rectangular ones, entirely on the local machine. Screen, SVG and PDF output are all derived from a single renderer-independent scene description, so exported figures reproduce the on-screen view exactly, and a black-and-white theme produces print-ready monochrome figures. A command-line mode allows batch conversion. PhyloGPU is written in Rust with the egui toolkit, ships as a single self-contained binary, and runs on Linux, macOS and Windows. Core algorithms are covered by automated tests, and exported figures were verified with an independent PDF renderer. PhyloGPU is intended for exploratory analysis, teaching and figure preparation rather than as a replacement for maximum-likelihood or Bayesian inference. Source code, examples and documentation are freely available under the MIT licence at https://github.com/sablokg/phylogpu.

Abstract: When taking pictures, there is often an unwanted foreground in the image, which may degrade the visual quality of the images. Separating unwanted foreground and background layers is an essential problem for enhancing image quality. However, image unwanted foreground detection and removal is a challenging process because of the ill-posed nature of the problem. Most previous works tackle this problem using multiple images or hand-crafted features. These methods are still quite limited in terms of generality and photo-realistic results. Detecting unwanted foregrounds automatically from the photos is highly desirable. This research presents an automated unwanted foreground detection technique based on deep learning. The model automatically detects the four types of unwanted foreground, including a fence, rain, fog, and shadow, from an image without degradation in image quality. The dataset used in this work consists of images taken from online repositories and self-captured. For shadow, the ISTD Dataset is used, a de-fencing dataset is used for fencing, and the Weather Image Recognition dataset images are used for fog and rain. Moreover, each type of unwanted foreground image is captured from the real world. To perform the unwanted foreground type detection, VGG-16 deep learning model is used. The experimental results indicate that the proposed system performs unwanted foreground detection with an accuracy of 97.13%. This research work can be used as first step towards the removal of unwanted foreground.

Abstract: Vision-based industrial safety monitoring increasingly integrates heterogeneous accident, personal protective equipment (PPE), and semantic evidence, yet preserving their reliability, uncertainty, and provenance throughout downstream decision support remains challenging. This study proposes an auditable downstream safety-intelligence framework that transforms frozen multimodal perception outputs into structured event evidence. Worker-centered temporal PPE observations are processed into Candidate and Confirmed operational evidence while accident-related uncertainty is preserved. In parallel, regulation-grounded retrieval maps the same event evidence to candidate regulatory context from a frozen Korean industrial-safety corpus, after which manager- and worker-oriented outputs are generated under distinct information boundaries and independently verified. On a fixed internal benchmark of 150 videos and 3,000 frames, the Full-clean configuration achieved 98.23% frame-level accuracy, 97.47% Macro-F1, and 98.67% video-level accuracy. Raw PPE-violation detection achieved an F1-score of 99.31% on the ground-truth normal-scene stratum. Controlled Legal-RAG regression reproduced 117/117 invariant retrievals, preserved query scope in 32/32 evidence-change cases, and produced the expected retrieval change in 1/1 query-changed case. All 234 notification–report links were audience- and event-consistent, while all 44 PPE semantic checks and 82 verifier findings satisfied predefined validation criteria. The framework supports provenance-preserving human review rather than autonomous safety or legal determination.

Abstract: Two assumptions dominate deep learning research on pneumonia detection from chest radiographs: that newer architectures outperform older ones, and that intra-dataset evaluation suffices to establish clinical utility. This study tested both. Nine architectures, spanning classical, efficient-scaling, and modern convolutional networks and Vision Transformers, plus a weighted soft-voting ensemble, were compared under a unified protocol on the pediatric Chest X-Ray Images dataset. None of the post-2020 architectures outperformed InceptionV3 (2016), which achieved the best individual accuracy (98.18%). The validation-weighted ensemble reached 98.52% accuracy and 99.22% pneumonia sensitivity on the internal test set; because its components were selected using test-set rankings, these values are upper-bound estimates. Bonferroni-corrected McNemar tests found no significant differences among the top-performing models. A cross-population external evaluation on an adult RSNA subset, which combines age, institutional, acquisition, and label shifts, revealed markedly different behavior: InceptionV3 retained 85.30% accuracy, whereas VGG16 fell to 59.60%, and the ensemble underperformed InceptionV3 alone. Ablation, Grad-CAM, and efficiency analyses complemented these findings. Architectural recency is not a proxy for performance, and external evaluation should be an architecture-selection criterion rather than an optional addendum.

Abstract: The conservation of industrial heritage has shifted from mere physical preservation to adaptive reuse; accurately distinguishing between historical components, later modifications, and other architectural elements is crucial for quantitative spatial analysis and renovation assessment. However, maintaining historic character often makes elements with different semantic attributes visually similar. Key challenges include subtle local differences between similar classes, insufficient coordination between local details and global context, and class imbalance causing intra-class dispersion and inter-class overlap. Using the Tianjin First Machine Tool Fac-tory as a case study, this study constructs a 14-class foreground semantic dataset and proposes Enhanced Mask2Former for Heritage Gardens (Hg-Mask2Former++). Based on Mask2Former-Swin-B, the model introduces LGFE to enhance multi-scale local–global representations and CDCC to optimize class relationships through class-balanced sampling and supervised contrastive learning. With only a 1.49% parameter increase, Hg-Mask2Former++ achieves 73.92% mIoU and 82.67% mDice, improving the baseline by 2.12 and 1.86 percentage points. Under identical experimental settings, its mIoU exceeds that of eight comparison models, including FastVim, TPSegformer, and ViT-P-based models proposed in 2026; the IoU of Industrial Building and Modern Building improves by about 20 percentage points. Further experiments reveal class-selective and condi-tion-dependent gains, indicating that scene-specific targeted optimization is more suitable than simply increasing network complexity.

Abstract: Chest radiography is the most common diagnostic X-ray examination, and paired images and reports are abundant in routine care, yet lesion-level spatial annotation remains scarce. This structured narrative review examines how radiology text has been used to supervise localization in chest radiographs: phrase grounding, weakly supervised abnormality localization, language-guided segmentation, grounded report generation, and multimodal large language models with spatial reasoning. Literature was drawn from a Web of Science seed search, citation chaining, and targeted searches through September 2026, and is organized along one axis, the granularity at which text binds to pixels, in five levels from study-level alignment to promptable or generative grounding. We show the axis is better read as two: train-time binding and inference-time spatial output come apart, and the systems that decouple them are the informative ones. Benchmarks such as MS-CXR and PadChest-GR made local evaluation realistic, and anatomy-aware resources such as PAXRay indicate that structural supervision improves grounding. Lesion masks now exist for the largest report corpus but are derived from those same reports, so model and reference share failure modes. We argue that anatomy-referenced evaluation, computable today, is the most tractable near-term correction, and that cross-study score comparison remains unsafe.

Abstract: Atlantic salmon aquaculture faces persistent welfare challenges from handling, confinement, disease and delousing, yet established welfare indicators typically require capture or invasive sampling and signal problems only after they have developed. Melanin-based pigmentation in salmonids is associated with stress responsiveness, but quantitative methods for measuring stress-associated fading of individual spots are lacking, and longitudinal photometry in repeatedly imaged individuals is vulnerable to between-session drift in illumination and camera settings. We present a computer vision framework that quantifies opercular melanin spots while explicitly controlling for image acquisition. Opercular regions were detected with YOLOv8, registered by projective transformation, and spots segmented with point-prompted fine-tuned SAM 2 model. Images were linearized and normalized to a common sensor exposure using EXIF metadata prior to feature extraction, and all effects were further adjusted for residual per-image exposure differences. Twenty-one salmon were imaged before and after 18 h confinement (676 matched spot pairs); plasma cortisol confirmed the challenge (W=14, p < 0.001). Eight control fish were imaged once as a reference group. Confinement was associated with increased spot grayscale intensity, reduced spot-to-background contrast and reduced spot area, with closely matching estimates on both opercula. Control fish were indistinguishable from pre-stress fish in contrast (p=0.28) but differed markedly from post-stress fish (d=1.04, p=0.002). Since control fish were imaged at a single timepoint, elapsed time cannot be excluded as a contributor. The framework provides a photometrically controlled, non-invasive basis for quantifying pigmentation dynamics in salmon.

Abstract: Manual annotation remains the source of ground truth for point cloud semantic classification, but its tools assume a contributor who can download billions of points and run desktop software. We present a browser-based framework that lets contributors classify remotely hosted datasets they neither own nor download. Points are labeled with a volumetric brush, and only the octree nodes the brush boundary crosses are fetched: regions fully inside are classified without being loaded, so the cost of a stroke follows its boundary rather than the points it encloses. Edits are held in a small ledger file, separate from the source data, that survives the discarding and reloading of points a streaming viewer performs and can be applied back to the source files at any time. Contributors receive disjoint regions as URLs and work asynchronously, returning ledgers that a project manager merges; classified subsets can also be exported as tiled LAS/LAZ. Class schemas are user-defined and travel with the edits. Where geometry alone is ambiguous, contributors label through pose-registered survey photographs. We demonstrate the framework on cultural heritage stonework, post-fire aerial LiDAR, and coral reef photogrammetry. It is open-source and deployed in production across OpenHeritage3D, the ALERTCalifornia Atlas Archive, and OpenOcean3D.

Abstract: Independent photogrammetric reconstructions of an object in different states require a compatible representation for interactive transitions. We present the ODD INN pipeline linking state-specific acquisition, alignment, Blender template fitting, appearance preparation and UV-carried deformation to coordinated geometry and material transitions in Unreal Engine. Its large-scale application pursued summer and winter capture across an entire open-air museum: ten major sessions generated more than 2 TB of mixed photographic, reconstruction-project and 3D data, with twenty multi-state objects selected from a larger collection. Architectural reconstructions, the ClayPot_01 preparation sequence and the playable Odmieniec demonstrator establish production and application outcomes; related assets are distributed commercially on Fab. Evaluation combines project component tests, production and engine demonstrations, and supplementary checks of the same technology. Recalculation reproduced all six fitting counts for three object pairs, whose outputs preserve source connectivity without exactly zero-area triangles. A Blender 5.2.0 test preserved base geometry and UV arrays through FBX export and reimport. These supplementary checks did not repeat fitting or Unreal execution. The contribution is an implemented acquisition-to-interaction pipeline with an inspectable geometry–appearance interface, demonstrated at museum-campaign and playable-environment scales.

Abstract: Multi-state photogrammetry records different conditions of the same physical object, but interactive use requires compatible geometry, paired appearance data and coordinated controls. We present the development and evaluation of the ODD INN system, linking offline fitting and asset preparation in Blender to geometry and material transitions in Unreal Engine 5. A summer–winter campaign aimed to scan an entire open-air museum; ten large sessions produced more than 2 TB of campaign data. Selected assets were integrated into Odmieniec, a playable environment with sliders, group controls, a freezing projectile, an expanding cold sphere and scene-wide transformations. The implementation uses one blend mask to coordinate geometric displacement, colour, normals and material properties. Evaluation combines integration tests, gameplay demonstrations, quality-assurance and rendering observations with supplementary source inspection and analysis of 199 frames from three recordings. These sequences show intermediate appearance, editable groups and repeated bidirectional changes; project reports document restoration and changes to at least five local and twenty global objects. The contribution is an implemented path from prepared observations to coordinated rendering and scoped interaction. The supplementary 2026 checks did not repeat Unreal execution. Geometric accuracy, restoration reliability and isolated transition cost remain unquantified.