Search Results (691)

A parking space system is a vital component of a smart transport management system, and it helps in managing parking spaces, reducing traffic congestion, and improving mobility. Deep learning helps in releasing smart parking systems. Despite their higher computational efficiency, lightweight convolutional neural networks (CNNs) may have limited feature representation capabilities and suffer from the vanishing gradient problem, which can reduce classification performance under challenging parking scenarios. To address this limitation, lightweight CNN models, such as MobileNetV2, EfficientNetB0, DenseNet, and ConvNeXt, are considered, with squeeze-and-excitation (SE) blocks incorporated to enhance channel-wise feature recalibration and stabilize gradients throughout the network. MobileNetV2 captures discriminative local features. EfficientNetB0 learns multi-scale semantic representations through compound scaling. DenseNet promotes hierarchical feature reuse. ConvNeXt extracts robust contextual features. The extracted deep features are subsequently reduced in dimensionality using Uniform Manifold Approximation and Projection (UMAP) while preserving the underlying manifold structure and removing redundant information. The reduced features are then classified by the ensembling of XGBoost, LightGBM, and support vector machine (SVM) classifiers, and their predictions are combined through a weighted stacking ensemble to exploit their complementary strengths and improve generalization. The proposed framework was experimentally evaluated on the PKLot, CNRPark and CNRPark + Ext datasets, achieving classification accuracies of 99.3%, 98.7% and 99.1%, respectively, demonstrating its effectiveness and robustness for real-world smart-city parking applications. Full article

Images captured under practical low-light conditions typically suffer from insufficient brightness, color distortion, noise, and blur, and the enhancement process itself may further introduce overexposed highlights. This paper presents HLCNet, an HVI-guided Large-Small Convolutional Cross-Branch Network for low-light restoration. RGB inputs are transformed into the HVI space so that chromatic and intensity information can be enhanced in two complementary branches. Each branch applies LSConv to couple broad illumination context with local structural modeling, followed by SE channel recalibration and an LCA-based encoder–decoder, while a soft overexposure constraint suppresses excessive responses without hard clipping. To ensure a controlled comparison, CIDNet is reproduced in the same Tesla T4 environment, whereas the published CIDNet results and other previously reported values are explicitly marked as external references. On LOL-Blur, HLCNet raises the PSNR of the reproduced CIDNet baseline from 26.5438 dB to 27.6260 dB, increases the SSIM from 0.8839 to 0.8863, and reduces the LPIPS from 0.1224 to 0.1056. On LOL-v2 Real and Synthetic, it attains 23.843 dB and 25.991 dB PSNR, respectively. Ablation, sensitivity, qualitative, and perceptual color-space analyses indicate that HLCNet is particularly effective for low-light images containing blur and weak structural details. Full article

Object detection remains a fundamental challenge in computer vision and plays a pivotal role in safety-critical medical applications, including surgical instrument recognition and operating-room workflow automation. This study presents a comprehensive comparative evaluation of five attention mechanisms—Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), Simple Attention Module (SimAM), and an enhanced multi-kernel Spatial Pyramid Pooling Fast module (SPPF+)—integrated into the YOLOv26n backbone, together with two neck-level attention variants (ECA-Neck and CBAM-Neck). A total of 16 model configurations were systematically investigated on a 22-class dental surgical instrument detection dataset under both default training settings and hyperparameter configurations optimized using the Optuna Tree-structured Parzen Estimator (TPE), enabling a rigorous full-factorial ablation study. Experimental results demonstrate that TPE-based hyperparameter optimization consistently enhances detection performance across all architectures. Among the evaluated models, CBAM-Opt achieved the highest detection accuracy, attaining an mAP@50 of 0.959 and an F1-score of 0.913, although the margins among the top optimized configurations fall within run-to-run variability. In contrast, Base-Opt delivered the strongest strict-localization capability with an mAP@50–95 of 0.800, highlighting the competitive performance of the baseline architecture when appropriately optimized. Notably, the parameter-free SimAM module exhibited the largest improvement following optimization ($\Delta$mAP@50 = +0.040), indicating a pronounced sensitivity to training configuration. Furthermore, neck-level attention integration achieved performance comparable to backbone-based attention, with ECA-Neck-Opt reaching an mAP@50 of 0.959, suggesting an effective alternative that preserves pretrained feature representations while maintaining high detection accuracy. Beyond performance evaluation, this work provides a unified benchmarking framework for attention mechanisms in medical object detection, accompanied by computational complexity analysis and practical architectural insights. The findings establish evidence-based guidelines for selecting attention modules in resource-aware surgical vision systems and contribute toward the development of more accurate and reliable computer-assisted clinical workflows. Full article

Recognizing facial emotions automatically from images/videos (FER) still represents a difficult problem for emotion computing, mainly due to variations in the face pose, lighting, occlusion, facial features, and expression intensity in the wild. Recent CNN–Transformer-based hybrid models like POSTER have leveraged local feature learning, landmark guidance, and global dependency modeling to achieve strong performance. Yet these methods give the main focus to spatial and contextual representations while not really going deep into adaptive channel-wise feature importance over multi-scale representations. As different feature channels represent emotions in varying degrees, it is likely that by treating all feature channels equally, one would limit the ability of the learned features to discriminate effectively. To overcome this weakness, this article presents a ResNet-18–Transformer landmark-guided module called SE-POSTER that fuses lightweight Squeeze-and-Excitation (SE) attention modules into the multi-scale feature pyramid of the baseline POSTER architecture. The proposed method carries out feature channel recalibration adaptively at the level of features before Transformer-based global attention modeling, thus allowing the network to focus on emotionally informative feature channels and suppress less relevant responses. The inclusion of SE attention in the network enhances fine, mid, and global levels of feature representations at a very low cost in terms of computation. On the basis of the RAF-DB, FERPlus, and AffectNet datasets, enormous experiments prove that the SE-POSTER framework proposed is capable of steadily boosting recognition accuracy relative to the baseline POSTER and several state-of-the-art FER methods. Especially, the proposed model delivers 92.78% accuracy on RAF-DB while it also shows better robustness and generalization capability under difficult real-world conditions. Moreover, additional ablation studies reveal that multi-level channel recalibration is effective in improving discriminative emotional feature learning. Full article

Korla fragrant pear disease and pest detection faces challenges such as significant object scale variation, multi-organ target confusion, and limited computational resources for real-time inference on edge devices. This study presents KFP-YOLO, a lightweight object detector based on YOLO26n, and constructs the Korla Fragrant Pear Disease and Pest Dataset (KFP-PDD), which covers leaves, fruits, and flowers and contains 14,092 original images, 22,735 annotated instances, and 11 healthy, disease, and pest categories. In KFP-YOLO, ADown modules are introduced into the backbone to reduce downsampling redundancy, selected C3k2 blocks are replaced with C3-PD modules combining partial convolution and squeeze-and-excitation attention, and selected feature-fusion nodes are redesigned as CFA modules incorporating coordinate attention. Under identical training and evaluation settings, KFP-YOLO reduces the number of parameters from 2.51 M to 1.93 M and the computational complexity from 5.79 to 4.42 GFLOPs, corresponding to reductions of 23.1% and 23.7%, respectively. On the KFP-PDD test set, KFP-YOLO achieves an mAP@0.5 of 0.9429 and an mAP@0.5:0.95 of 0.7434, compared with 0.9458 and 0.7579 for the YOLO26n baseline. Its inference speed reaches 278.97 FPS on the Jetson AGX Orin platform. In addition, evaluation on an independent external test set containing 1052 images and 3377 annotated instances yields a Precision of 0.8814, a Recall of 0.9077, an mAP@0.5 of 0.9325, and an mAP@0.5:0.95 of 0.7322. These results indicate that KFP-YOLO provides a favorable trade-off between detection accuracy, model complexity, and edge inference efficiency, although its detection accuracy is slightly lower than that of the baseline. The proposed model therefore provides a lightweight candidate for further validation in continuous orchard monitoring applications. Full article

When marine chillers operate under complex marine conditions, they suffer from severe cross-equipment feature distribution shifts, scarce labeled fault samples in the target domain, industrial vibration noise mixed in sensor signals, and difficulties in accurately identifying subtle faults with varying severity levels. To tackle these issues, this paper improves upon the domain difference perception network (DDPN) and proposes a dual-hybrid attention feature discriminant domain-Adversarial network (DAFDAN) to realize intelligent fault diagnosis across different equipment and working conditions under few-shot scenarios. The proposed method constructs a dual-branch feature encoder consisting of a source domain compressor and a target domain extender to accommodate the distinct sensor dimensions of two heterogeneous chiller types. A hybrid attention module is formed by integrating squeeze-and-excitation efficient channel attention (SE-ECA, a module for screening channel-wise features) and spatial attention, which adaptively amplifies time-series features sensitive to faults and suppresses irrelevant noise. Residual connections (shortcut paths in deep neural networks to mitigate the vanishing gradient problem during deep-layer training) are introduced to optimize feature transmission. A dual-layer domain alignment framework is built with gradient reversal layers and maximum mean discrepancy (MMD). Combined with adversarial training (a training paradigm that learns domain-agnostic features through a game between a feature extractor and a domain discriminator), the framework achieves joint optimization of implicit feature confusion and explicit distance constraints. Meanwhile, a five-stage progressive training strategy is designed, which activates multiple loss functions, including weighted cross-entropy, mean square error (MSE), binary cross-entropy (BCE), and Kullback–Leibler (KL) divergence stage by stage. Class weighting and early stopping strategies are adopted to alleviate sample imbalance and model overfitting. In this paper, the public ASHRAE RP-1043 centrifugal chiller dataset is used as the source domain, and time-series measurement data collected from a self-developed laboratory marine screw chiller serves as the target domain. Verification experiments are carried out covering one normal steady-state operating condition and 15 gradient faults falling into five major categories with different severity degrees. Results from ablation experiments (controlled-variable comparative experiments that quantify the independent contribution of each component by comparing model performance with or without a specific module/loss), multi-algorithm comparisons, and confusion matrix visualization demonstrate that the cross-domain fault diagnosis accuracy of the proposed DAFDAN approaches is 100%, outperforming mainstream transfer learning algorithms such as support vector machine (SVM), deep neural network (DNN), MMD, correlation alignment (CORAL), and domain-adversarial neural network (DANN). Multiple ablation experiments verify that the three core components—hybrid attention, adversarial training, and semi-supervised learning—jointly boost the model’s diagnosis accuracy and operational stability. The loss curves of the complete five-stage training process converge smoothly. The confusion matrix reveals zero misjudgments and zero false alarms across all 16 refined operating states, enabling precise identification of subtle incipient faults of all severity levels. This study proves that DAFDAN can effectively address the pain points of few-shot cross-equipment fault diagnosis for marine chillers and provides a reliable algorithmic reference for the intelligent operation and maintenance of ship refrigeration equipment. Full article

To address the challenges of analyzing vibration signals, overlapping fault characteristic frequency bands, and difficulties in accurately distinguishing similar faults, this study proposes a diagnostic classification scheme for GIS disconnect switch mechanical fault detection. The scheme integrates multi-scale RGB image representation with a Squeeze-and-Excitation Convolutional Neural Network Attention (SE-CNN-Attention) mechanism. First, continuous wavelet transform is employed to decouple the one-dimensional vibration signal into low-, medium-, and high-frequency physical bands, which are then mapped to the red, green, and blue channels to generate a 64 × 64 × 3 RGB time–frequency feature image, enabling three-dimensional encoding of time, frequency, and energy intensity. Based on this approach, a SE-CNN-Attention fusion model is developed. The two-dimensional CNN automatically extracts region-specific features from the time–frequency images. Concurrently, the squeeze-and-excitation (SE) attention mechanism adaptively assigns weights to enhance fault-sensitive frequency bands and suppress noise interference. The model adopts global average pooling instead of traditional fully connected layers, combined with Dropout regularization and a Softmax classifier, to achieve efficient and robust fault classification. Experiments are conducted under four operating conditions with a total of 19,800 sample groups collected. The findings indicate that the proposed approach attains an overall classification accuracy exceeding 94%, outperforming traditional support vector machine (SVM), baseline convolutional neural network (CNN), Vision Transformer (ViT), and ResNet-18 models. The SE-Attention module significantly enhances the model’s focus on key frequency band features while accelerating convergence speed and improving classification stability. This method provides reliable technical support for intelligent maintenance and fault warning of GIS disconnect switch equipment and holds significant engineering application value. Full article

Tunnel lining defects pose significant threats to structural integrity and operational safety. Traditional image processing and machine learning methods often suffer from limited accuracy and poor generalization under complex backgrounds. To address these limitations, this study proposes a lightweight intelligent recognition method based on You Only Look Once version 11 nano (YOLOv11n) for Ground Penetrating Radar (GPR) images of tunnel linings. The backbone is replaced with Mobile Network Version 3 (MobileNetV3) to reduce parameters and Floating Point Operations (FLOPs), while depthwise separable convolution and a streamlined Compressed 2-Stage Fused-Lite (C2f-Lite) structure are integrated into the Neck to further decrease computational overhead. Channel mapping layers are employed to ensure smooth feature transfer, and selective use of Squeeze-and-Excitation (SE) attention and Hard-Swish (H-swish) activation balances detection accuracy with efficiency. Evaluated on a low-power mobile workstation acting as an edge-precursor proxy platform, experimental results demonstrate that the improved YOLOv11n_MobileNetV3 model achieves high accuracy with a mean Average Precision (mAP) at 0.5 of 94.4% and mAP@0.5:0.95 of 62.4%, low computational cost of 4.7 Giga Floating Point Operations (GFLOPs), and fast inference speed of 45 Frames Per Second (FPS). Comparative analysis further confirms its superior balance of detection performance and efficiency over YOLO version 5 (YOLOv5) and YOLO version 8 (YOLOv8) baselines. The proposed approach provides a highly optimized, edge-oriented engineering solution for real-time tunnel lining defect inspection, establishing strong structural and theoretical feasibility for future deployment in embedded systems. Full article

The present work proposes EGT-UNet, an innovative loss optimization mechanism based on Evolutionary Game Theory (EGT) aimed at pediatric panoramic tooth segmentations. In the absence of a large set of high-quality images in the current literature, we developed a novel pediatric panoramic image database, including 1269 images from Ordu University Faculty of Dentistry. The dataset includes subjects aged 3–14, where 67% are males and 33% are females. The annotation of all images was done at the pixel level by a professional orthodontist with a two-fold verification process. In order to prevent data leakage, the dataset was split at the subject level into training/validation (85%) and independent test (15%) sets. Data in the training set were divided into folds for five-fold cross-validation. From a structural point of view, EGT-UNet is an improved version of U-Net, equipped with the following modules: Squeeze-and-Excitation blocks, Attention Gates, and a dilated convolutional module mimicking Atrous Spatial Pyramid Pooling. The main novelty of the proposed approach involves the application of the dynamic change in loss weights. Specifically, in our work, the fusion of three losses—Dice, Focal Tversky and Boundary—was dynamically tuned via replicator dynamics using task-specific fitness functions. The difference with traditional loss-weighting methods is that in the latter case, fixed weights are applied. To evaluate the effect of each component of the EGT-UNet, we conducted an ablation study of six architectures varying in terms of hybrid loss function and dynamic/static weight tuning. Using the independent test dataset, we observed a similar performance level of all models with Dice scores ~0.93. Our best model, called “EGT Aggressive”, achieved Dice = 0.931 ± 0.044, IoU = 0.873 ± 0.066, and Boundary Dice = 0.631 ± 0.064. Importantly, this model demonstrated a statistically significant superiority over the baseline network according to region-related metrics and boundary metrics (Wilcoxon p < 0.05). In computational studies, we revealed an increased model robustness when dealing with highly complex mixed dentition images. Full article

With the rapid development of remote sensing technology, high-resolution satellite imagery has been increasingly applied to land resource monitoring, urban planning, and environmental assessment. Automatically assigning semantic labels to remote sensing image patches remains a fundamental challenge due to pronounced intra-class variation and high inter-class visual similarity. To address the trade-off between model capacity and limited labeled data, this paper proposes a remote sensing image scene classification framework based on an improved EfficientNetV2-S architecture. The proposed model integrates a Squeeze-and-Excitation (SE) channel attention module between the final 1 × 1 expansion convolution and the Global Average Pooling layer, where it functions as a late-stage channel gating mechanism that adaptively recalibrates channel-wise responses, though its accuracy benefit is seed-sensitive rather than consistently reproducible at the current dataset scale. A two-stage optimization strategy was evaluated, comprising a fully unfrozen supervised baseline followed by a pseudo-label semi-supervised fine-tuning stage utilizing a strict confidence threshold ($\tau =0.90$). Evaluated on a 10-class subset of the public NWPU-RESISC45 benchmark, the purely supervised SE-EfficientNetV2-S delivers 98.71% independent test accuracy, matching or exceeding the much larger ResNet50 (98.50%, 24.1 M parameters) despite using only 20.4 M parameters. Multi-seed variance analysis further reveals that semi-supervised fine-tuning yields a small test-set improvement for the No-SE configuration that is consistent in sign across all three seeds (+0.46 pp mean) but not statistically significant at this sample size, and an even smaller, likewise non-significant gain for the SE-augmented model (+0.08 pp), suggesting that channel gating moderates pseudo-label effectiveness in small-data regimes. Full article

6D object pose estimation from RGB-D data is a core problem in robotic perception. Representative approaches such as DenseFusion fuse per-pixel RGB and depth features, yet they still face challenges under heavy occlusion and in cluttered backgrounds. This paper revisits the DenseFusion architecture from the perspective of feature enhancement and global context modeling and proposes an enhanced network named RAFusion, which strengthens feature representation through two complementary modules. First, Squeeze-and-Excitation (SE) blocks are integrated into the RGB feature extractor to adaptively reweight channel responses, enhancing salient appearance cues while suppressing noise. Second, a RealFormer-based block is introduced before global pooling, applying residual self-attention to the fused RGB-D features for more effective global context modeling. Experiments on the LINEMOD and Occlusion LINEMOD datasets show that RAFusion improves the DenseFusion baseline and achieves competitive performance compared with representative 6D pose estimation methods. Ablation studies and computational analysis further support the effectiveness of the proposed modules. Full article

To address the challenges of low detection accuracy and weak generalization in identifying mining-induced ground fissures from UAV imagery, caused by their slender and discontinuous morphology, complex background clutter, and multi-scale surface features, this paper proposes an enhanced deep semantic segmentation model, SAG-DeepLabV3+ (with Spatial Vision Transformer, Attention mechanisms, and Adaptive Gated Fusion). Specifically, to enhance global context modeling and fine boundary delineation, we introduce a Spatial Vision Transformer (SVT) branch within the Atrous Spatial Pyramid Pooling (ASPP) module. We further employ a dual attention mechanism, sequentially combining Squeeze-and-Excitation (SE) and a Convolutional Block Attention Module (CBAM), for progressive channel and spatial feature refinement. Moreover, an Adaptive Gated Fusion (AGF) module is designed to dynamically optimize the fusion of multi-level decoder features. Experiments on a dedicated UAV-based mining fissure dataset comprising 1280 annotated images show that SAG-DeepLabV3+ achieves a state-of-the-art mean Intersection over Union (mIoU) of 79.52% (with Xception backbone) and 79.19% (with lightweight MobileNetV2 backbone), surpassing DeepLabV3+, U-Net, and PSPNet by a significant margin. Furthermore, by leveraging transfer learning (pre-training on the public CrackVision12K dataset and fine-tuning on our mining fissure dataset), the model’s mIoU is further elevated to 82.04%, demonstrating superior generalization capability. The proposed SAG-DeepLabV3+ effectively balances high accuracy with operational efficiency, fulfilling the potential demand for lightweight automated fissure monitoring under resource-limited field deployments, and lays a foundation for subsequent real-time on-site deployment verification. Full article

Low-light color image enhancement remains a challenging task for vision-based decision systems, which must simultaneously address illumination correction, noise suppression, contrast recovery, and color preservation from a single degraded observation. This study proposes ResGASP-GAN, a GAN-based low-light image enhancement framework built around a residual-output generator that integrates batch-size-independent normalization, multi-scale contextual aggregation, and channel-wise feature recalibration within a conditional adversarial setting. Group Normalization is integrated into the proposed generator to reduce dependence on batch statistics during small-batch training, while a Squeeze-and-Excitation (SE) module adaptively enables channel-wise feature recalibration and helps preserve structural and chromatic information. The proposed generator uses reflection-padded convolutions to reduce boundary artifacts, and it features a multi-scale bottleneck composed of dilated residual blocks and Atrous Spatial Pyramid Pooling to capture spatially varying illumination patterns. The model is optimized using a compound objective that combines an adversarial term with an ${\ell }_1$ reconstruction loss, balancing perceptual realism with pixel-level fidelity. Experimental evaluation employed the LOL-v1, LOL-v2-Real, and LOL-v2-Synthetic datasets using reference-based metrics: PSNR, SSIM, and LPIPS. No-reference perceptual metrics were also used, including NIQE and BRISQUE. The results indicate that the proposed method achieves competitive structural similarity and visual image quality on LOL-v2-Real, competitive reconstruction performance on LOL-v1, and good generalization on LOL-v2-Synthetic, with the second-best metrics of PSNR = 22.30 dB, SSIM = 0.9154, and LPIPS = 0.1022 among the reported methods on this last dataset. In contrast, using no-reference metrics, this study achieved very good results, with the lowest BRISQUE of 10.4540 and a competitive NIQE of 4.4396, providing high-quality visual–perceptual reconstruction. Overall, the proposed architecture provides a competitive GAN-based alternative for low-light image enhancement, combining residual connections, multi-scale contextual modeling, and channel-wise feature refinement. This architecture provides the lowest inference time over all discussed models, which is highly required in real-time outdoor applications such as robot navigation and mapping. Full article

Vibration-based mechanical fault diagnosis has become a critical research area, mostly driven by the need to improve equipment reliability and reduce unplanned downtime in industrial settings. Time–Frequency Networks (TFNs) have shown strong potential here, combining interpretable time–frequency transformations with deep learning classifiers in a single framework. This work reproduces the original TFN model from the recent literature and evaluates its noise robustness under additive Gaussian noise (10 dB, 0 dB, $-5$ dB SNR) and impulsive noise at the same levels, across five architectures: Backbone CNN, Random CNN, TFN-Chirplet, TFN-Morlet, and a squeeze-and-excitation attention CNN baseline. The evaluation protocol corrects two methodological issues identified during peer review of an earlier version of this work—window-level data leakage between train and test splits, and selection of the best-performing training epoch rather than a fixed final-epoch result—both of which are shown to materially affect reported outcomes. Under the corrected protocol, TFN-Morlet remains the most noise-robust architecture, with only a 19.09% accuracy drop from clean to $-5$ dB AWGN, approximately 15.5 percentage points better than Backbone CNN under the same conditions; an architectural anomaly reported in the earlier version of this study, in which mild noise appeared to improve an unconstrained CNN’s accuracy, was not reproduced under the corrected protocol and is shown to be an artifact of the original methodological issues. Per-class analysis and multi-model confusion matrices further reveal that misclassifications under severe noise are dominated by confusion between the same defect severity at different fault locations, rather than between different severities at the same location as previously reported. These results indicate that time–frequency-aware convolutional kernels improve both classification accuracy and noise resistance under rigorous, leakage-free evaluation, and that this robustness is not replicated by a generic attention mechanism alone. Full article

Robust monitoring of weld pool dynamics is critical for automated arc welding; however, single-modality sensors are frequently constrained by severe optical interference and high-frequency environmental noise. To address these limitations, we propose M2WPR-Net, a novel multimodal framework that synergizes visual and acoustic signals for simultaneous weld width regression and physical quality classification. The architecture employs a dual-stream ResNet50 backbone to process heterogeneous sensory data. Specifically, the visual stream utilizes a Convolutional Block Attention Module (CBAM) to suppress intense arc glare and localize the weld pool. Concurrently, the acoustic stream transforms 1D audio sequences into 2D Gramian Angular Summation Field (GASF) textures, which are subsequently refined by Squeeze-and-Excitation (SE) networks to isolate target frequency channels. A central contribution of this study is a bidirectional cross-modal attention mechanism based on Query–Key–Value (Q-K-V) matrix operations. Overcoming the shortcomings of static feature concatenation, this module dynamically aligns the modalities, enabling acoustic cues to guide visual feature extraction and vice versa, thereby mitigating information bottlenecks. Optimized via a joint multi-task loss function, the proposed M2WPR-Net significantly outperforms existing single-modal and conventional fusion baselines. Experimental results demonstrate that the network achieves a Mean Absolute Error (MAE) of 0.18 mm for width prediction and a 93.5% accuracy in penetration state classification, confirming its resilience and practical applicability in complex industrial welding environments. Full article