Search Results (2,257)

Drainage pipe defect detection is essential for maintaining the normal operation of urban infrastructure. In recent years, deep learning-based object detection methods have provided an effective technical solution for drainage pipe defect recognition. Among them, YOLO-series models have demonstrated strong potential in visual detection tasks due to their end-to-end architecture and high inference efficiency. However, directly applying baseline YOLO models may still face challenges such as limited detection accuracy, relatively high model complexity, and insufficient adaptability for lightweight deployment scenarios. To address these issues, this paper proposes a lightweight drainage pipe defect detection method based on an improved YOLO11 network. Rather than treating detection enhancement and model compression as two separate procedures, the proposed method integrates feature enhancement, adaptive pruning, and distillation-based recovery into a unified lightweight detection framework. Specifically, an improved SimAM attention mechanism is introduced into the backbone and integrated with the C3k2 module to construct the C3K2_SWS module, aiming to enhance the representation capability of critical defect features. In the neck network, a focused diffusion pyramid network with a dimension-aware selective fusion structure, termed FDPN-DASI, is designed to strengthen multi-scale feature interactions. In addition, an adaptive-threshold focal loss (ATFL) is introduced to improve the learning capability for hard samples. For efficient deployment, the LAMP pruning algorithm is further improved, and an entropy-guided entropy-adaptive magnitude-based pruning method (EA-LAMP) is proposed to enable adaptive allocation of pruning ratios across different network layers. Moreover, BCKD knowledge distillation is applied after pruning to mitigate the accuracy degradation caused by model compression. Experimental results indicate that the proposed lightweight YOLO11-SFA+EA+BCKD framework achieves a precision of 92.4%, a recall of 88.5%, and an mAP50 of 93.3%, while maintaining a compact model size of 1.6 M parameters and 4.5 G FLOPs. Compared with the baseline model, the proposed method improves precision, recall, and mAP50 by 5.9%, 5.0%, and 4.7%, respectively, while reducing the number of parameters, FLOPs, and model size by 1.0 M, 1.8 G, and 2.1 M, respectively. These results suggest that the proposed framework can improve detection performance while reducing model complexity under the current experimental setting, indicating its potential for lightweight drainage pipe defect detection tasks. Full article

Grape caliber distributions are critical for packing, grading, yield estimation, and post-harvest logistics. However, estimating reliable caliber histograms from single images remains challenging due to occlusion and dense bunch structure. This work presents a two-stage monocular pipeline that integrates instance segmentation, geometry-aware representations, residual quantity correction, and structured histogram prediction. In the first stage, a YOLO-based model detects grape instances and a calibration object, enabling the construction of geometry-aware auxiliary channels and a segmentation-derived counting prior. In the second stage, these representations are used to estimate total grape count and caliber distributions. Results show that RGBDT consistently outperforms RGB, indicating that geometry-aware cues improve both histogram fidelity and counting accuracy. The framework achieves stable performance under realistic conditions while maintaining low runtime, supporting practical deployment in agricultural environments. Full article

Short-term traffic forecasting is a critical component of intelligent transportation systems. While deep learning architectures for this task have proliferated rapidly, the sensor-level data characteristics—zero-value prevalence, distributional heterogeneity, and cross-sensor correlation structure—that drive architecture-specific failure modes remain insufficiently understood, and their implications for evidence-based model selection in real deployments have not been systematically addressed. This study addresses that question through a sensor-network diagnostic framework applied to the METR-LA dataset (Metropolitan Los Angeles; 207 inductive loop detectors, 5-min resolution). The framework integrates systematic characterization of sensor data properties, a controlled benchmark of four representative architectures—Transformer, Spatio-Temporal Graph Convolutional Network (STGCN), Diffusion Convolutional Recurrent Neural Network (DCRNN), and Gated Temporal Convolutional Network (Gated TCN)—under a unified 12→3 prediction setting, and a novel per-sensor regression analysis that quantitatively links zero-value ratios to model-specific prediction errors across all 207 sensors. Building on these findings, this study further proposes Graph-Enhanced Transformer (GETFormer), a lightweight hybrid architecture that augments the Transformer with a single-hop Graph Convolutional Network (GCN) layer and a gated residual fusion module. The diagnostic findings and condition-dependent model-selection guidelines provide an empirically grounded foundation for principled hybrid architecture development in urban traffic sensing. Full article

Floods remain among the most devastating natural disasters globally, disproportionately impacting low-resource regions where real-time flood forecasting is constrained by limited computational infrastructure and the scarcity of fine-resolution predictive models. Although state-of-the-art global climate models achieve high predictive accuracy, their scale and computational complexity restrict their applicability in localized and resource-constrained settings. This study proposes a deep climate model distillation framework that transfers knowledge from a high-capacity Fourier Neural Operator (FNO)-based global climate model inspired by FourCastNet into lightweight, regionally adaptive student networks suitable for edge deployment. The framework combines climate variables, satellite observations, and hydrological measurements to improve localized flood prediction. Knowledge transfer is achieved through a multi-objective distillation strategy that combines supervised learning, soft-target alignment, and intermediate feature matching. Experimental evaluation across multiple flood-prone regions in Sub-Saharan Africa and South Asia shows that the distilled student model achieves an average classification accuracy of 0.89, an AUC of 0.91, and an F1-score of 0.88, retaining approximately 96.7% of the teacher model’s predictive performance. In continuous discharge estimation, the model attains a mean absolute error of 0.17, RMSE of 0.24, and an R² score of 0.85. The proposed distillation approach yields an 8× reduction in inference latency and over a 20× reduction in model size, enabling real-time execution on low-power edge devices such as the Raspberry Pi 4 and NVIDIA Jetson Nano. The student model further demonstrates robust regional and temporal generalization, with limited performance degradation in unseen geographic areas and during extreme flood years. Full article

Edge vision systems require convolutional neural networks (CNNs) that preserve recognition accuracy under strict storage, computation, and latency constraints. Although ResNet18 is a compact residual backbone, direct deployment on resource-constrained devices remains costly, whereas simple channel reduction weakens representation capacity. This study aims to build a deployable ResNet18-based classifier that reduces model complexity while recovering the accuracy lost during compression. We propose a lightweight framework that combines global channel scaling, a re-parameterized attention residual block, and teacher–student knowledge distillation. The proposed block uses multi-branch convolution and squeeze-and-excitation attention during training, then folds the linear branches into a single 3-by-3 convolution for inference. Experiments on CIFAR-100 show that the final model reduces parameters from 11.220 M to 2.841 M, retains comparable Top-1 accuracy (0.7579 vs. 0.7606), improves Top-5 accuracy (0.9340 vs. 0.9253), and reduces graphics processing unit (GPU) batch inference latency from 3.279 ms to 2.161 ms. Deployment on PYNQ-Z2 verifies the complete camera-based CPU-side inference workflow, with an average end-to-end latency of 421.467 ms/frame. The results indicate that residual topology preservation, re-parameterized feature enhancement, and distillation form a practical route for edge-oriented lightweight CNN deployment. Full article

Determining the 3D position of a vehicle from a 2D image plays a key role in video surveillance, autonomous driving, and spatial localization. However, localization accuracy can significantly degrade in conditions of incomplete or synthetic measurement noise and keypoint jitter. In this paper, we propose a new iterative 3D position estimation algorithm (KGA). This algorithm includes geometric correction and calibration steps for converting from 2D to 3D coordinates; trajectory prediction and correction using a Kalman filter; and adaptive tuning of the filter parameters using the Goldschmidt algorithm. Experiments confirm that KGA outperforms the standard (FK) and modified (MFK) Kalman filters in accuracy and convergence speed, demonstrating robustness to various camera angles and noise levels. The novelty of this approach lies in the integration of the Goldschmidt algorithm into the Kalman filter to create an adaptation mechanism that dynamically adjusts the measurement noise covariance based on instantaneous innovation magnitude. Unlike end-to-end deep learning trackers or nonlinear filters (EKF/UKF), KGA is designed as a lightweight post-processing stage that can be seamlessly integrated into existing detection pipelines while maintaining the low computational footprint required for UAV-based edge deployment. The algorithm is of practical value for computer vision systems requiring accurate and robust tracking under varying observational conditions, with current implementation suitable for offline or buffered processing, and clear pathways to real-time deployment through code optimization. The algorithm is of practical value for computer vision systems requiring accurate and robust tracking under varying observational conditions. Full article

Precision agriculture increasingly relies on machine learning applied to high-resolution data acquired by unmanned aerial vehicles (UAVs) to support crop monitoring, stress detection, and yield forecasting. This survey presents a structured review of machine learning methods for UAV-enabled precision agriculture and organizes over 100 peer-reviewed studies within a unified four-dimensional taxonomy defined by sensing modality, data type, model family, and analytical task. The taxonomy enables systematic comparison across RGB, multispectral, hyperspectral, LiDAR, and IoT data sources and across classical machine learning, deep learning, hybrid sequential models, and emerging transformer-based architectures. We analyze how modeling choices interact with data characteristics to influence robustness, cross-environment generalization, computational efficiency, and deployment feasibility on UAV and edge platforms. Recurring challenges include limited labeled data, domain shift across seasons and fields, multimodal heterogeneity, occlusion, and real-time processing constraints. We identify emerging research directions, including data-efficient learning, representation-level multimodal fusion, domain adaptation, lightweight architectures for embedded deployment, and uncertainty aware decision support. By formalizing the landscape through a unified taxonomy, this survey provides a foundation for designing scalable, robust, and deployable machine learning systems for next-generation precision agriculture. Full article

Recent advancements in computer vision, particularly in detection, segmentation, and classification, have significantly impacted various domains. However, these advancements are still strongly tied to RGB-based systems, which are insufficient for applications in industries such as waste sorting, pharmaceuticals, and defence, where material characterization beyond shape or visible colour is necessary. Hyperspectral (HS) imaging captures spatial and spectral information for each pixel and therefore offers a promising route for material-level classification. This study evaluates the potential of combining HS imaging with deep learning for plastic material classification. The work includes: (i) the design of an experimental setup with a HS line-scan camera, conveyor, and controlled illumination; (ii) the construction of an object-disjoint dataset of HDPE, PET, PP, and PS samples with semi-automated mask generation and Raman spectroscopy-based labelling; and (iii) the development of P1CH, a lightweight pixel-wise 1D convolutional hyperspectral classifier. On object-disjoint test images, P1CH achieved 97.44% all-pixel accuracy. A boundary sensitivity analysis, reported separately because semi-automated labels are uncertain at material/background interfaces, yielded 99.94% accuracy after excluding a pre-defined two-pixel border band. Additional ablation, baseline, and robustness analyses show that the proposed pixel-wise spectral approach is effective for small fragments, visually similar plastics, and overlapping materials, while black or very dark plastics remain challenging under the present camera and illumination configuration. Full article

The proliferation of Unmanned Aerial Vehicles (UAVs) in various applications has created a pressing need for robust and efficient communication systems. Fifth-generation (5G) networks can support UAV connectivity through high bandwidth and low-latency communication; however, rapid three-dimensional UAV mobility creates handover-management challenges that can increase signalling overhead, service interruption, and Quality of Service (QoS) degradation. This paper presents an integrated framework that combines LSTM-based multi-UAV trajectory prediction with proactive handover optimization using an Advantage Actor–Critic (A2C) Deep Reinforcement Learning (DRL) agent. The LSTM predictor is evaluated on a real-world UAV trajectory dataset and reports a root mean square error (RMSE) of 4.37 m over a 5 s prediction horizon after conversion to a local East–North–Up coordinate frame. A lightweight simulation-level coordination mechanism is included to reduce simultaneous target-cell contention among multiple UAVs; it is not claimed as a new standardized 3GPP signalling procedure. Handover performance is evaluated by replaying 180 held-out flight trajectories in a controlled 5G simulation across ten independent random seeds. Under these stated assumptions, the proposed framework achieves a handover success rate of $94.2\pm 0.8$%, an average SINR of $15.8\pm 0.2$ dB, a handover delay of $45.2\pm 1.1$ ms, and a handover frequency of $0.85\pm 0.05$ HOs/min, outperforming the tuned 3GPP A3, reactive SINR, and CASH baselines in the reported simulation results (Wilcoxon signed-rank test, $p<0.01$, Bonferroni-corrected). The experimental setup is described in detail to support methodological transparency and facilitate future replication, but the handover results should be interpreted as simulation-based evidence rather than live-network validation. Full article

Urban acoustic environments and performing-arts spaces strongly influence speech communication quality, acoustic comfort, and public wellbeing, particularly in noise-exposed shared environments such as transport hubs, campuses, healthcare spaces, public service facilities, music-education settings, and rehearsal or performance-related spaces. To address speech–noise separation in low signal-to-noise ratio and acoustically complex scenarios, this study proposes a lightweight two-stage deep learning framework termed LCGRU–Wave-SkipConvNet. In the preprocessing stage, a Lightweight Convolutional Gated Recurrent Unit (LCGRU) model is employed to achieve preliminary separation of target speech and background noise by capturing both spatial and temporal acoustic features. In the post-processing stage, a Wave-SkipConvNet model is introduced to further suppress residual noise and enhance speech quality. Experimental results demonstrate that the proposed framework achieves superior performance under different signal-to-noise ratios, sound-source angles, and target angle errors. For example, in the preprocessing stage, the LCGRU model achieved a perceptual evaluation of speech quality (PESQ) score of 2.64 at source angles between 0° and 30°, outperforming the convolutional neural network-long short-term memory (CNN-LSTM) model by 1.17. In the post-processing stage, the Wave-SkipConvNet model achieved higher short-time objective intelligibility (STOI) and segmental signal-to-noise ratio (segSNR) values than the comparison models under different SNR conditions. The proposed framework provides an effective and deployment-oriented AI solution for improving speech accessibility and acoustic comfort in urban acoustic environments and performing-arts spaces. Beyond speech enhancement, it offers practical potential for supporting healthier, more inclusive, and more equitable acoustic environments in noise-sensitive public and educational spaces. It should be noted that this study focuses on the objective acoustic environment and signal-level speech enhancement, rather than subjective soundscape perception, musical perception, or human perceptual evaluation. Full article

Accurate fault classification of electro-hydrostatic actuators (EHAs) remains challenging because multivariate fault signals contain local transient variations, inter-variable coupling, and dynamic temporal dependencies that are difficult to capture simultaneously using a single model. To address this problem, this paper proposes a multi-view temporal feature collaborative heterogeneous ensemble learning model (MTF-HEM) for EHA multivariate time series fault classification. MTF-HEM integrates a representative subsequence-guided time series forest (RSG-TSF), XGBoost, and a lightweight LSTM to extract local morphological, global statistical, and temporal dependency features, respectively. The outputs of these heterogeneous base learners are fused using a bootstrap-driven out-of-bag probability binning stacking (BOPB-stacking) strategy. The proposed method was evaluated on an AMESim-based simulated EHA plunger pump fault dataset containing one normal condition and six fault conditions. Under the present simulation setting, MTF-HEM achieved an accuracy of 99.52% and outperformed the tested deep time series classification models, ensemble models, and individual base learners. These results suggest that multi-view heterogeneous feature fusion can improve the classification of simulated EHA fault time series and provide a methodological reference for intelligent actuator fault diagnosis. However, the current validation is based on data generated from a single AMESim simulation model, and further evaluation on real EHA systems is needed to assess the practical applicability and generalizability of the proposed approach. Full article

Fish observation is a key component of marine ecological monitoring and is valuable for understanding ecological processes and fish population dynamics. In practical applications, observation equipment is often constrained by limited memory and computational resources, making it difficult to deploy visual detection models with large parameter counts and high computational complexity. Under limited computational resources, existing deep-learning-based fish detection models struggle to balance detection accuracy, model lightweighting, and real-time edge deployment. To address this issue, a lightweight GEM-YOLOv8n model based on YOLOv8n is proposed for fish detection. For high model complexity, insufficient feature representation, and limited bounding box regression accuracy in complex underwater observation scenarios, the model replaces C2f modules and some Conv modules with GhostC2f and GhostConv, introduces the EMA attention mechanism, and adopts MPDIoU loss instead of CIoU. Experimental results show that, compared with YOLOv8n, GEM-YOLOv8n improves Precision, Recall, mAP50, and mAP50–95 by 0.53%, 2.16%, 0.52%, and 0.34% while reducing parameters and FLOPs by 48.0% and 39.5%. These results demonstrate that the proposed model improves detection performance while reducing model complexity. Tests on Jetson Xavier NX demonstrate real-time performance and deployment feasibility, providing a lightweight deployment solution for resource-constrained underwater fish detection. Full article

The aim in single-image deraining tasks is to remove rain streaks from degraded images while preserving scene structures and fine details. However, existing deep learning-based methods often face a trade-off between restoration quality and computational efficiency, and many models struggle to capture hierarchical information effectively under complex rain conditions. To address these limitations, we propose a lightweight cross-gated hierarchical transformer for image deraining. The proposed network adopts a five-stage encoder–decoder architecture with Multi-head Feature-wise Attention (MFA) to efficiently model channel-wise dependencies while reducing the computational burden associated with conventional self-attention. In addition, an Enhanced Gated Depthwise Feed-Forward Network (EGDFN) is introduced to obtain refined feature representations with improved efficiency, and a Cross-Level Feature Refinement (CLFR) module is designed to enhance information exchange between corresponding encoder and decoder stages, thereby strengthening hierarchical feature integration and preserving structural details. The network is trained using a single SSIM-based loss, which enhances the structural fidelity of the restored results. Extensive experiments on four synthetic datasets, two real-world datasets, and a downstream semantic segmentation benchmark demonstrate that the proposed method consistently achieves strong restoration performance, producing cleaner outputs with sharper details and improved effectiveness for subsequent vision tasks. Full article

Human sitting posture recognition is critical for smart seating, ergonomic monitoring, and healthcare systems. However, existing deep learning approaches typically rely on highly complex network architectures that are computationally expensive, hindering their lightweight deployment on edge devices. Furthermore, current methods frequently struggle with indistinct boundaries among multi-class postures and are highly prone to overfitting when constrained by small-sample pressure sensor datasets. To bridge this gap, this paper proposes a novel, lightweight posture recognition framework specifically tailored for pressure distribution maps. First, sitting pressure data is collected using a thin-film pressure array sensor and uniformly mapped into an [M × N] image representation, establishing an effective sample format for Convolutional Neural Network (CNN) inputs. Second, as our primary architectural contribution, we fundamentally optimize the classic LeNet-5 network to enhance complex feature representation without inflating model complexity. Specifically, the depth of the convolutional layers is increased with a progressively increasing channel configuration. Batch Normalization (BN) is introduced to accelerate convergence and ensure training stability, while a Dropout mechanism is embedded within the fully connected layers to strictly penalize overfitting under small-sample constraints. These architectural improvements are synergistically combined with targeted data augmentation strategies—including random translation, rotation, and intensity perturbation—to further strengthen the model’s generalization capability. Experimental results demonstrate that the proposed method achieves a classification accuracy of 95.5% in a five-class sitting posture recognition task, significantly outperforming baseline models such as the traditional LeNet-5, AlexNet-Lite, and VGG-Small. The findings indicate that this approach achieves an optimal balance among recognition accuracy, training stability, and low model complexity, providing a robust algorithmic baseline and proof-of-concept for smart healthcare perception systems, paving the way for future large-scale subject-independent validation. Full article

The early and accurate recognition of fruit tree pests and diseases is essential for safeguarding fruit yield, quality, and sustainable agricultural production. Conventional manual inspection methods are inadequate for meeting the demands of continuous, objective, and real-time monitoring in large-scale orchards. Following the framework of “model evolution–key challenges–edge-intelligent deployment,” this review systematically summarizes advances in deep learning-based recognition of fruit tree pests and diseases, and compares the effectiveness and limitations of representative methods from the perspectives of data complexity, model generalization and robustness, real-time inference, cross-modal fusion, and trustworthy diagnosis. Existing studies indicate that CNNs, attention mechanisms, Transformers, multimodal fusion, and lightweight networks have promoted the transition of fruit tree pest and disease recognition from image classification to object detection, lesion segmentation, and edge deployment; however, sample scarcity, class imbalance, insufficient cross-domain generalization, black-box decision-making, energy constraints, and long-term robustness remain major bottlenecks for field application. Future research should focus on open orchard environments and develop data-efficient, interpretable, low-power, and continuously updatable edge-intelligent recognition systems, thereby advancing precision agriculture and smart orchards. Full article