Search Results (69)

Hyperspectral imaging (HSI) has potential for forensic entomology, but its practical use is limited by manual region-of-interest (ROI) delineation before spectral extraction. This step is time-consuming, operator-dependent, and difficult to standardize across insect species and developmental stages. Here, we developed an automated individual-level ROI-to-spectrum workflow for HSI analysis of forensically important insects. The dataset included 63 hyperspectral images and 1868 manually annotated insect individuals, covering larvae, pupae, and adults. The proposed Hyperspectral Imaging Fully Convolutional Network (HSI-FCN) segmented insect body regions from three-band pseudo-RGB images, back-projected the predicted masks to the original HSI data cubes, generated individual-level ROIs, and extracted full-band mean spectra. On an independent test set containing 204 insect individuals, HSI-FCN achieved mean Dice and intersection over union (IoU) values of 0.9079 and 0.8328, respectively, and showed the best overall performance among representative segmentation models. All test individuals were successfully matched with their corresponding manual ROIs. Spectra extracted from automated ROIs were highly consistent with manual ROI spectra, with a mean spectral angle mapper of 3.06° and a Pearson correlation coefficient of 0.9956. These results show that the proposed workflow can replace manual ROI delineation with a reproducible preprocessing step for insect HSI analysis, supporting standardized spectral extraction and future applications in forensic entomology. Full article

In response to Saudi Arabia’s amended White Land Fees Law, which imposes charges of up to 10% of land value on undeveloped urban plots, this study presents TerraVision, an intelligent framework for large-scale White Land detection and urban land monitoring using high-resolution satellite imagery and deep learning. The proposed framework aims to support sustainable urban development by enabling municipalities and planners to identify underutilized urban land, improve land-use efficiency, and support evidence-based planning decisions. Satellite imagery was acquired through the Esri ArcGIS platform at a spatial resolution ranging from 0.31 to 0.34 m per pixel. The Riyadh study area was divided into 1317 geographic tiles, of which 80 tiles covering approximately 180 km² were manually annotated to construct the training and evaluation dataset. Ten segmentation models representing four architectural families were evaluated, including encoder–decoder networks, transformer-based architectures, YOLO segmentation models, and the zero-shot Segment Anything Model 3 (SAM3). Six fine-tuned semantic segmentation models achieved Intersection over Union (IoU) scores between 0.94 and 0.96 on the held-out test set, with SegFormer achieving the highest performance at an IoU of 0.9563. A post-inference geoprocessing pipeline was developed to reconstruct city-scale prediction maps, estimate neighborhood-level White Land availability, and export results into GIS- and web-compatible formats. The framework was further integrated into a bilingual (Arabic and English) decision-support dashboard that enables visualization and spatial analysis of vacant land distribution. The results demonstrate that semantic segmentation models provide an accurate solution for monitoring undeveloped urban land that scales to city-wide inference across Riyadh, and can support preliminary screening for strategic urban planning and sustainable city development initiatives in Riyadh. Full article

Introduction: Fish species represent 25% of all vertebrates across the globe and are one of the most threatened animal groups. At least 40% of the fish fauna occurs in rivers for part of their life cycle. European rivers are home to more than 600 fish species, while also being some of the most impaired and altered ecosystems. Objective: The objective was to assess the vulnerability of freshwater fish communities in European river basins. Methodology: Using RivTool and the CCM2 database, we developed the River Restoration Units (R2Us), a set of spatial units that takes into account river network functioning and allows a higher spatial discretisation than river basins. We developed RivFish, a database about the presence of native freshwater-dependent fish in 1556 Europeans river basins. For this, we collected data from 77 references and validated synonyms and scientific names for 667 species. We used the latest International Union for Conservation of Nature (IUCN) Red List assessment to define species distributions in European rivers. After intersecting with the R2U layer, we curated and validated species names and spatial occurrence using RivFish. To map the vulnerability of freshwater fish communities, we used the Habitats Directive (HD) and the IUCN datasets. These consider a distinct number of species and assess conservation status differently: the HD evaluated 165 species, while the IUCN evaluated 516 species. The HD data allowed calculating the composite indicator of Conservation Status, whereas the IUCN data enabled calculating the vulnerability index. Results: Both ana-lyses show higher richness in central Europe, particularly in the Danube basin. Spatially, both highlight southern Europe as the area where fish communities have the highest vulnerability. However, the HD analysis also indicates the Danube and the western Atlantic basins as having high vulnerability. The IUCN analysis shows the Anatolian and Mediterranean biogeographical regions as those with the highest vulnerability values. Conclusions: Southern Europe’s higher vulnerability is likely associated with restricted distribution ranges and high levels of endemicity in Mediterranean fish communities. Overall, these findings improve current knowledge and show that input data may be key to effort allocation towards the management and conservation of European freshwater fish communities. Full article

Background/Objectives: In laparoscopic and robot-assisted surgery, bleeding may rapidly impair operative-field readability and procedural safety. In the broader Robin Heart teleoperation framework, interpretation of such events is relevant not only for scene understanding but also as a potential prerequisite for future safety-oriented supervisory functions under communication-degraded conditions. The aim of this study was to assess whether a deep learning model for blood segmentation could provide outputs suitable for preliminary image-level temporal characterization of visible blood-region behavior in laparoscopic video. Methods: A U-Net-based binary blood-segmentation model was implemented in-house in PyTorch and evaluated on three paired image–mask datasets: a simulated bleeding dataset prepared under controlled laboratory conditions, an internal operative laparoscopic dataset, and an external-domain subset derived from the public GynSurg dataset. Segmentation performance was assessed using 5-fold cross-validation and reported using the Dice coefficient and Intersection over Union (IoU). Training dynamics were analyzed using training and validation loss and Dice curves. Additional baseline comparisons were performed on the internal operative dataset using U-Net++ and DeepLabV3+. Temporal analysis was performed on selected video fragments, including a low-motion reference sequence without active bleeding progression, internal bleeding-related sequences, and external-domain sequences, using mask-derived descriptors and auxiliary optical-flow-based motion descriptors computed after camera-motion compensation within the detected blood-related ROI. Results: In 5-fold cross-validation, the U-Net-based model achieved Dice coefficient and IoU values of 0.915 ± 0.012 and 0.851 ± 0.019 on the simulated dataset, 0.856 ± 0.013 and 0.756 ± 0.025 on the internal operative dataset, and 0.707 ± 0.053 and 0.570 ± 0.056 on the external-domain GynSurg subset, respectively. On the internal operative dataset, the proposed model performed comparably to U-Net++ and slightly above DeepLabV3+ under the same cross-validation protocol. The temporal descriptor set differentiated low-motion reference behavior, more spatially coherent progression, rapid coherent expansion, and dynamic or motion-active progression profiles. Peak dA/dt reflected abrupt visible blood-area expansion, temporal IoU described mask stability over time, and optical-flow-based descriptors provided additional information on local motion activity within the detected blood-related ROI. Conclusions: The results support the feasibility of combining deep-learning-based blood segmentation with temporal and optical-flow-based descriptors for exploratory image-level characterization of visible blood-region behavior in laparoscopic video. Within the Robin Heart development pathway, such descriptors may, in the future, serve as candidate components of image-analysis support modules for safety-oriented teleoperative scenarios. At this stage, they should be interpreted as exploratory image-derived indicators rather than clinically validated markers of bleeding severity. Full article

Landslides are among the most destructive geological hazards in Malaysia, especially in mountainous and forested areas. Unmanned aerial vehicle (UAV) imagery offers high spatial resolution and flexible data capture, but deep learning performance is highly sensitive to training hyperparameters. In this study, the TransU-Net model for UAV-based landslide detection was adopted and a systematic ablation study on learning-rate and epoch settings using a coarse-to-fine tuning strategy. The Berembun Forest Reserve dataset was first used to determine the optimal training configuration. Then, the optimised configuration was tested on multiple UAV sub-datasets in the CAS Landslide dataset to evaluate performance stability under different terrain properties and spatial resolutions. The optimised configuration yielded the best F1-score (0.9598) and IoU of 0.9507 on the Berembun Forest Reserve dataset, and consistently high F1-scores across the evaluated CAS Landslide sub-datasets. Qualitative visualisation analysis also revealed good spatial correspondence between the predicted segmentation masks and the ground-truth annotations. Variations in Intersection over Union (IoU) values were mainly associated with boundary delineation uncertainty rather than severe misclassification. Overall, the results show that the performance of UAV-based landslide segmentation can improve by systematic hyperparameter tuning, and the optimised TransU-Net configuration under the evaluated terrain conditions yields promising results. Full article

Murals are not merely independent visual artworks. Rather, they are an integral part of architectural heritage, directly attached to buildings’ structural elements, such as brick walls and vaults. However, murals are susceptible to various building-related types of damage, including structural cracks and moisture-induced peeling, due to long-term exposure to environmental factors and geological changes. As the progressive deterioration of these murals hastens the loss of mural value, professional assessment and restoration are urgently required. To tackle the issues of low efficiency in traditional structural damage detection and the absence of predictable repair plans, this paper presents a semi-automatic building-mural protection solution that integrates morphological assessment of mural deterioration with computer vision technology. This study establishes an image prediction system that integrates intelligent damage identification with virtual restoration. First, employing the PaddleSeg deep learning framework and the DeepLabv3 semantic segmentation model, this study used existing mural damage datasets to build a recognition model. The model allows for intelligent identification and labeling of multiple damage types. Subsequently, relying on the ComfyUI platform, Stable Diffusion was used to construct a virtual restoration model. LoRA (low-rank adaptation) technology was introduced to fine-tune the model specifically for the mural style, thus enhancing the directivity and accuracy of virtual restoration. Finally, by applying the results of the recognition model to the virtual restoration model, this study built an integrated system for mural damage diagnosis and virtual restoration. The results show that the damage recognition model achieved a mean intersection over union (mIoU) of 47.8% and a pixel accuracy of 77.97% on the test set, validating the feasibility of using semantic segmentation for mural damage detection. This study presents an integrated workflow framework integrating automatic damage identification and intelligent repair. As an expert-assisted tool, this framework shows application potential for preliminary exploration of mural disease diagnosis and virtual restoration plans, providing technical references for the digital protection of cultural heritage. Full article

This paper presents a reproducible system innovation for processing Digital Imaging and Communications in Medicine (DICOM) mammography images based on pixel-monotonic dynamic range management and engineering-verifiable intensity transformations. Standard DICOM conversion schemes to 8-bit representation often result in irreversible luminance-range compression, locality-dependent contrast distortions, and reduced robustness of deep learning models. The proposed framework preserves the physical consistency of the Modality LUT and photometric polarity, performs breast-aware robust Winsor normalization, and applies strictly monotonic global tone mapping while preserving the 16-bit depth of the training data. System validation was performed using architecture-independent metrics. Compared to standard processing, the median value of normalized mutual information increased from 0.878 to 0.892, the effective number of bits increased from 7.88 to 10.11 (+2.25), the representation entropy increased by 1.42 bits, and the clipping rate was reduced to almost zero. Experiments with the Faster R-CNN detector showed stable or improved calcification localization at Intersection over Union (IoU) ≥ 0.5 under controlled augmentation conditions. The results confirm that pixel-monotonic dynamic range control provides a reproducible, engineering-verifiable basis for AI-based mammography analysis within the evaluated dataset and experimental setting. Full article

Background/Objectives: Intraoral radiographs acquired using photostimulable phosphor (PSP) plates are inherently susceptible to a wide spectrum of artifacts that can compromise diagnostic reliability and lead to unnecessary repeat exposures. Although structured taxonomies describing these artifacts have been proposed, automated methods capable of detecting and localizing multiple artifact types at the pixel level remain limited, particularly under realistic multi-class conditions. In this study, we address the problem of fine-grained, multi-class PSP artifact segmentation by systematically evaluating a deep learning-based framework and establishing a realistic baseline for this inherently challenging task. Methods: A retrospective, multi-center dataset comprising 1497 intraoral PSP radiographs (bitewing and periapical) collected from three institutions was analyzed. Pixel-level annotations were generated by expert oral and maxillofacial radiologists according to a standardized taxonomy consisting of four major artifact groups and 29 artifact classes, together with a background class. A 2D nnU-Net v2 architecture was employed as a baseline segmentation model. Model development was performed using 5-fold cross-validation, and performance was evaluated on an independent test set using Dice coefficient, Intersection over Union (IoU), Precision, and Recall. Results: Across all classes, the model achieved a mean Dice score of 0.0894 ± 0.0084 in cross-validation and 0.0952 on the independent test set, reflecting the intrinsic complexity of the task. Class-wise analysis revealed substantial variability, with higher performance in larger and visually distinctive artifacts, whereas small-scale, low-contrast, and underrepresented classes exhibited markedly reduced performance. Notably, several artifact categories were absent from the training data, resulting in a zero-shot scenario that directly constrained model generalization. Furthermore, segmentation performance demonstrated a strong dependency on class frequency, measured in terms of pixel distribution, underscoring the impact of severe class imbalance. Group-based evaluation showed relatively higher performance for pre-exposure and exposure-related artifacts compared to post-exposure and scanner-related categories. Conclusions: These findings demonstrate that large-scale, multi-class pixel-level segmentation of PSP artifacts represents a fundamentally challenging problem shaped by the combined effects of class imbalance, small object size, heterogeneous artifact morphology, and incomplete training representation. While the proposed framework confirms the feasibility of automated artifact localization, its current performance suggests greater immediate value as a quality control or screening support tool rather than a fully autonomous diagnostic system. By providing a comprehensive baseline and systematic analysis, this study establishes a benchmark for future research and highlights the critical need for imbalance-aware learning strategies, hierarchical modeling, and data-centric approaches to advance this field. Full article

There have been significant advancements in object detection using extensive labelled datasets. However, existing learning-based approaches remain constrained in industrial environments, primarily due to the limited diversity in training datasets; the lack of generalisation of close-set detectors to unseen asset categories; and the inherent spatial and geometric complexity of mechanical, electrical, and plumbing (MEP) assets. To address this challenge, we propose a new approach that leverages pre-trained vision language models and close-set object detectors to detect unseen MEP assets using unlabelled data. Experimental results reveal the superior performance of Grounding DINO using Swin B transformer in open-vocabulary MEP asset detection, achieving the mean intersection over union (mIoU) of 0.6586 for valve detection and 0.4883 for pump detection. In addition, the combination of Grounding DINO (Swin B) and YOLOv8 outperforms other configurations in MEP asset detection, attaining the highest performance for both valve detection, with mean average precision at IoU = 0.5 (mAP50) of 0.928 and mean average precision over IoU threshold from 0.5 to 0.95 (mAP50:95) of 0.889, and pump detection, with corresponding values of 0.778 and 0.662, respectively. The quantitative and qualitative results of our approach were evaluated against fine-tuned Grounding DINO and fully supervised close-set object detectors. Full article

Background: Bone Mineral Density (BMD) plays a crucial role in diagnosing osteoporosis, and early detection is essential to preventing complications such as osteoporotic fractures. However, access to dual-energy X-ray absorptiometry (DXA) screening remains limited in many healthcare settings. Objective: This study presents a fully automated artificial intelligence pipeline for BMD prediction from lumbar spine radiographs to enable opportunistic osteoporosis screening. Methods: The proposed system integrates automatic vertebral segmentation and a machine learning-based regression model for BMD prediction. A YOLO-based instance segmentation model was trained to automatically segment four lumbar vertebrae, achieving a high Intersection over Union (IoU) of 0.9. Radiomic features were extracted from the segmented vertebrae to capture advanced image characteristics and combined with clinical features from 2875 female patients. An eXtreme Gradient Boosting (XGBoost) regressor was trained to provide opportunistic BMD estimation. Results: The model achieved a mean absolute percentage error (MAPE) of 6% for BMD prediction. A classification model built from segmented vertebrae distinguished between osteoporosis, osteopenia, and normal bone with approximately 90% accuracy. Strong agreement between predicted and ground-truth BMD values was confirmed using Pearson correlation coefficient and Bland–Altman analysis. Conclusions: The proposed fully automated system demonstrates strong agreement with DXA measurements and potential for opportunistic osteoporosis screening in settings with limited DXA access. Further validation and refinement are needed to achieve clinical-grade precision for diagnostic applications. Full article

Fractures are a leading cause of morbidity and mortality in Thoroughbred racehorses, posing a significant threat to their welfare and careers. This study introduces a deep learning model specifically designed to facilitate fracture detection in equine athletes. By leveraging extensive training on human fracture data and refining the model with equine imaging, it highlights the transformative potential of transfer learning across species and medical contexts. This approach is not limited to equine fractures but could be adapted for use in detecting injuries or conditions in other veterinary species and even human healthcare applications. A comprehensive databank of radiographs, sourced from public archives and equine hospitals, was curated to encompass diverse conditions (fracture and non-fracture), ensuring robust pattern recognition. The architecture integrates a Vision Transformer for global context modelling with a ResNet backbone and loss function to optimize local feature extraction and cross-species adaptability. The pipeline achieved 96.7% accuracy for modality classification, 97.2% accuracy for projection recognition, and fracture localization intersection over union values of 0.71–0.84 across equine datasets. This work bridges advancements in human and veterinary medicine, opening pathways for AI-driven solutions that extend beyond fractures, fostering improved diagnostic precision and broader applications across species (felines, canines, etc.). By integrating advanced imaging techniques with AI, this study aims to set a foundation for more comprehensive and versatile health monitoring systems. Full article

Early detection of breast cancer commonly relies on imaging technologies such as ultrasound, mammography and MRI. Among these, breast ultrasound is widely used by radiologists to identify and assess lesions. In this study, we developed image segmentation techniques and multiclass classification artificial intelligence (AI) tools based on pretrained models to segment lesions and detect breast cancer. The proposed workflow includes both the development of segmentation models and development of a series of classification models to classify ultrasound images as normal, benign or malignant. The pretrained models were trained and evaluated on the Breast Ultrasound Images (BUSI) dataset, a publicly available collection of grayscale breast ultrasound images with corresponding expert-annotated masks. For segmentation, images and ground-truth masks were used to pretrained encoder (ResNet18, EfficientNet-B0 and MobileNetV2)–decoder (U-Net, U-Net++ and DeepLabV3) models, including the DeepLabV3 architecture integrated with a Frequency-Domain Feature Enhancement Module (FEM). The proposed FEM improves spatial and spectral feature representations using Discrete Fourier Transform (DFT), GroupNorm, dropout regularization and adaptive fusion. For classification, each image was assigned a label (normal, benign or malignant). Optuna, an open-source software framework, was used for hyperparameter optimization and for the testing of various pretrained models to determine the best encoder–decoder segmentation architecture. Five different pretrained models (ResNet18, DenseNet121, InceptionV3, MobielNetV3 and GoogleNet) were optimized for multiclass classification. DeepLabV3 outperformed other segmentation architectures, with consistent performance across training, validation and test images, with Dice Similarity Coefficient (DSC, a metric describing the overlap between predicted and true lesion regions) values of 0.87, 0.80 and 0.83 on training, validation and test sets, respectively. ResNet18:DeepLabV3 achieved an Intersection over Union (IoU) score of 0.78 during training, while ResNet18:U-Net++ achieved the best Dice coefficient (0.83) and IoU (0.71) and area under the curve (AUC, 0.91) scores on the test (unseen) dataset when compared to other models. However, the proposed Resnet18: FrequencyAwareDeepLabV3 (FADeepLabV3) achieved a DSC of 0.85 and an IoU of 0.72 on the test dataset, demonstrating improvements over standard DeepLabV3. Notably, the frequency-domain enhancement substantially improved the AUC from 0.90 to 0.98, indicating enhanced prediction confidence and clinical reliability. For classification, ResNet18 produced an F1 score—a measure combining precision and recall—of 0.95 and an accuracy of 0.90 on the training dataset, while InceptionV3 performed best on the test dataset, with an F1 score of 0.75 and accuracy of 0.83. We demonstrate a comprehensive approach to automate the segmentation and multiclass classification of breast cancer ultrasound images into benign, malignant or normal transfer learning models on an imbalanced ultrasound image dataset. Full article

The Qinba Mountains in China span six provinces, characterized by a large population, rugged terrain, steep peaks, deep valleys, and scarce flat land, making large-scale agricultural development challenging. Terraced fields serve as the core cropland type in this region, playing a vital role in preventing soil erosion on sloping farmland and expanding agricultural production space. They also function as a crucial medium for sustaining the ecosystem services of mountainous areas. As a transitional zone between China’s northern and southern climates and a vital ecological barrier, the Qinba Mountains’ terraced ecosystems have undergone significant spatial changes over the past two decades due to compound factors including the Grain-for-Green Program, urban expansion, and population outflow. However, current large-scale, long-term, high-resolution monitoring studies of terraced fields in this region still face technical bottlenecks. On one hand, traditional remote sensing interpretation methods rely on manually designed features, making them ill-suited for the complex scenarios of fragmented, multi-scale distribution, and terrain shadow interference in Qinba terraced fields. On the other hand, the lack of high-resolution historical imagery means that low-resolution data suffers from insufficient accuracy and spatial detail for capturing dynamic changes in terraced fields. This study aims to fill the technical gap in detailed dynamic monitoring of terraced fields in the Qinba Mountains. By creating image tiles from Landsat-8 satellite imagery collected between 2017 and 2020, it employs three deep learning semantic segmentation models—DeepLabV3 based on ResNet-34, U-Net, and PSPNet deep learning semantic segmentation models. Through optimization strategies such as data augmentation and transfer learning, the study achieves 15-m-resolution remote sensing interpretation of terraced field information in the Qinba Mountains from 2000 to 2020. Comparative results revealed DeepLabV3 demonstrated significant advantages in identifying terraced field types: Mean Pixel Accuracy (MPA) reached 79.42%, Intersection over Union (IoU) was 77.26%, F1 score attained 80.98, and Kappa coefficient reached 0.7148—all outperforming U-Net and PSPNet models. The model’s accuracy is not uniform but is instead highly contingent on the topographic context. The model excels in environments that are archetypal for mid-altitudes with moderately steep slopes. Based on it we create a set of tiles integrating multi-source data from RBG and DEM. The fusion model, which incorporates DEM-derived topographic data, demonstrates improvement across these aspects. Dynamic monitoring based on the optimal model indicates that terraced fields in the Qinba Mountains expanded between 2000 and 2020: the total area was 57.834 km² in 2000, and by 2020, this had increased to 63,742 km², representing an approximate growth rate of 8.36%. Sichuan, Gansu, and Shaanxi provinces contributed the majority of this expansion, accounting for 71% of the newly added terraced fields. Over the 20-year period, the center of gravity of terraced fields shifted upward. The area of terraced fields above 500 m in elevation increased, while that below 500 m decreased. Terraced fields surrounding urban areas declined, and mountainous slopes at higher elevations became the primary source of newly constructed terraces. This study not only establishes a technical paradigm for the refined monitoring of terraced field resources in mountainous regions but also provides critical data support and theoretical foundations for implementing sustainable land development in the Qinba Mountains. It holds significant practical value for advancing regional sustainable development. Full article

Lung cancer remains the leading cause of cancer-related deaths worldwide, with over 2.4 million new diagnoses in 2022. Early diagnosis remains challenging due to the non-specificity of symptoms, often resulting in late-stage detection. Although 2-D and 3-D medical imaging, particularly computed tomography (CT), is widely used for detecting lung cancer, it is associated with manual segmentation, which remains time-consuming and user-dependent. This study proposes LuCa as an innovative 2.5-D deep learning model for lung cancer delineation, which combines the benefits of 2-D segmentation with 3-D volume delineation. The main novelty of LuCa is focused on its pipeline, specifically designed to be of clinical use, in order to guarantee the usability of the method. LuCa employs a U-Net architecture for segmentation, followed by a post-image-processing step for 3-D tumor volume delineation and false-positive correction. The method was trained and evaluated using the “NSCLC-Radiomics” database, comprising CT images of 422 non-small cell lung cancer patients, with clinical manual tumor annotations as ground truth. The model achieved strong performance, with high dice coefficients (87 ± 12%), intersection over union (81 ± 17%), sensitivity (84 ± 16%), and positive predictive value (94 ± 10%) on the test set. Performance was particularly high for larger tumors, reflecting the ability of the model to delineate more visible lesions accurately. Statistical analysis confirmed the high correlation and minimal error between predicted and ground truth tumor volumes. The results highlight the potential of the 2.5-D approach to improve clinical efficiency by enabling accurate tumor segmentation with reduced computational cost, compared to traditional 3-D methods. Future research will focus on assessing the use of LuCa as real-time clinical decision support, particularly for assessing tumors during treatment. Full article

Underwater wireless optical communication (UWOC) is a key technology in ocean resource development, and its link stability is often limited by the difficulty of optical alignment in complex underwater environments. In response to this difficulty, this study has focused on improving the Real-Time Detection Transformer v2 (RT-DETRv2) model. We have improved the underwater light source detection model by collaboratively designing a lightweight backbone network and deformable convolution, constructing a cross-stage local attention mechanism to reduce the number of network parameters, and introducing geometrically adaptive convolution kernels that dynamically adjust the distribution of sampling points, enhance the representation of spot-deformation features, and improve positioning accuracy under optical interference. To verify the effectiveness of the model, we have constructed an underwater light-emitting diode (LED) light-spot detection dataset containing 11,390 images was constructed, covering a transmission distance of 15–40 m, a ±45° deflection angle, and three different light-intensity conditions (noon, evening, and late night). Experiments show that the improved model achieves an average precision at an intersection-over-union threshold of 0.50 (AP50) value of 97.4% on the test set, which is 12.7% higher than the benchmark model. The UWOC system built based on the improved model achieves zero-bit-error-rate communication within a distance of 30 m after assisted alignment (an initial lateral offset angle of 0°–60°), and the bit-error rate remains stable in the 10⁻⁷–10⁻⁶ range at a distance of 40 m, which is three orders of magnitude lower than the traditional Remotely Operated Vehicle (ROV) underwater optical communication system (a bit-error rate of 10⁻⁶–10⁻³), verifying the strong adaptability of the improved model to complex underwater environments. Full article