Aiming at the problems of large-scale span, extreme aspect ratio and extremely low pixel proportion of micro-cracks on the surface of civil engineering structures, this paper proposes an improved object detection algorithm based on YOLOv8. Based on YOLOv8, the algorithm first integrates the Convolutional Block Attention Module (CBAM) into the key feature extraction stage of the backbone network. Through feature weighting in both channel and spatial dimensions, it effectively suppresses noise interference in the concrete or pavement background and enhances the sensitivity to crack textures. Secondly, the Slicing Aided Hyper Inference (SAHI) strategy is introduced in the inference stage. Through overlapping slice processing and result fusion of high-resolution images, the problem of micro-crack feature loss caused by direct scaling of high-resolution images is fundamentally solved. Experimental results show that the model performs significantly better than the original YOLOv8 and other mainstream models on the crack detection datasets, with the m AP@0.5 increased from the original 55.2% to 71.2%, which verifies the effectiveness and practicability of the method in the field of building structure health monitoring.
Driver behavior classification is a critical component of intelligent transportation systems, enabling proactive road safety interventions and personalized driver assistance. This paper presents a comparative evaluation of five supervised machine learning algorithms — Naive Bayes, K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest, and Gradient Boosting — applied to the task of classifying three driving behaviors: Normal, Drowsy, and Aggressive. Features are extracted from multi-sensor vehicle data including GPS speed, three-axis accelerometer readings, gyroscope signals, lane deviation measurements, steering entropy, and brake frequency, inspired by the publicly available UAH-DriveSet benchmark. A dataset of 2,100 labeled instances is constructed with deliberate class overlap to simulate real-world ambiguity. Following standard preprocessing and 70/12.5/17.5 train/validation/test split, each model is evaluated on accuracy, precision, recall, F1-score, and five-fold cross-validation accuracy. The Naive Bayes classifier achieves the highest test accuracy of 95.48% and F1-score of 95.47%, demonstrating that carefully engineered sensor features can yield strong classification performance even with lightweight probabilistic models. SVM, Random Forest, and Gradient Boosting each achieve 95.24% accuracy, while KNN trails at 94.52%. Feature importance analysis identifies jerk mean, speed mean, and lane deviation standard deviation as the most discriminative signals. This study confirms that machine learning combined with sensor fusion can effectively support real-time driver monitoring systems.
Accurate prediction of temperature and humidity within grain storage facilities is essential for ensuring long-term food safety and reducing post-harvest losses. Existing approaches based on standard Long Short-Term Memory (LSTM) networks or Sequence-to-Sequence LSTM (Seq2Seq-LSTM) architectures often fail to capture the complex multi-scale temporal dependencies inherent in grain depot microclimate data. In this paper, we propose a hybrid deep learning model that integrates Convolutional Neural Networks (CNN), Bidirectional LSTM (BiLSTM), and a Multi-Head Attention mechanism (CNN-BiLSTM-Attention). The CNN component extracts local temporal patterns, the BiLSTM captures both forward and backward long-range dependencies, and the Multi-Head Attention layer dynamically focuses on the most informative time steps. We evaluate the proposed model on a real-world dataset collected from a grain storage depot in Northwest China spanning two full years of hourly records. Results demonstrate that the proposed model consistently outperforms SVR, LSTM, and Seq2Seq-LSTM baselines across prediction horizons of 6 h, 12 h, and 24 h. For 24-hour temperature prediction, the model achieves MAE of 0.89°C and coefficient of determination of 0.9713, representing 35.5% and 4.2% improvements over Seq2Seq-LSTM.
To address the issues of lacking unified representation for multi-source heterogeneous data in oil and gas field development, poor interpretability of traditional data-driven methods, and the difficulty in structuring expert experience, a causal knowledge graph construction method integrating data-driven and large language models (LLMs) is proposed. This method first learns the causal skeleton from structured data using causal discovery algorithms (PC) and then utilizes LLMs to extract domain knowledge from engineering documents, performing semantic completion for edges with undetermined directions. On this basis, an LLM-based agent is introduced to validate the physical consistency and compliance of the causal graph according to safety constraints and engineering rules, and an incremental update mechanism is designed to adapt to dynamic production environments. Experiments on actual oil field datasets demonstrate that the causal graphs generated by this method are highly consistent with expert annotations, achieving an F1 score of 0.90, significantly outperforming traditional methods in metrics such as true positive rate and precision, effectively realizing the synergy between data patterns and domain knowledge, and providing a new pathway for explainable intelligent analysis in oil and gas fields.
To address the low efficiency, high alignment error, and insufficient positioning reliability of single-sensor manual alignment in bulk cement truck loading, this paper proposes a circular target detection and localization method based on color space conversion, morphological filtering, and multi-sensor concurrent fusion. The proposed method first performs Gaussian preprocessing on industrial camera images to suppress high-frequency noise caused by dust, vibration, and illumination fluctuation. The filtered image is then converted from the RGB color space to the HSV color space, where calibrated dual red thresholds are used to obtain a binary mask of the circular port marker. Morphological erosion, dilation, opening, and closing are introduced to remove isolated noise, suppress edge burrs, and repair small discontinuities in the segmented region. On this basis, the FindContours algorithm extracts connected components, the largest valid contour is selected by geometric constraints, and minimum enclosing circle fitting is used to calculate the pixel center and radius of the target. To overcome the lack of absolute depth in monocular vision, an Arduino-based ultrasonic ranging module is further integrated, and a Python multiprocessing shared-memory mechanism is designed to decouple high-frame-rate image processing from low-baud-rate serial communication. Experimental results show that the proposed method effectively suppresses complex industrial background interference, outputs stable center coordinates, radius, and depth information, and accurately triggers the stop decision when the visual and ultrasonic thresholds are simultaneously satisfied, thereby improving the robustness and real-time performance of automated loading alignment.
Lightweight single-image super-resolution is important for edge-side imaging systems where reconstruction accuracy, model size, memory movement, and inference latency must be considered together. Although many compact SR networks reduce floating-point operations through channel distillation or depth wise-style operators, fewer FLOPs do not necessarily lead to faster execution on real hardware because intermediate feature access may still be expensive. In addition, when the distillation branch is overly compressed, shallow features may be forwarded with insufficient nonlinear transformation, which weakens the recovery of complex textures. To address these problems, this paper proposes PCF-IMDN, a compact feature-distillation network based on partial-convolution feature extraction. In the proposed block, spatial convolution is applied only to a selected subset of channels, and a following 1×1 pointwise convolution is used to fuse cross-channel information. This design reduces redundant spatial filtering while preserving feature interaction. Moreover, a lightweight 1×1 transformation is inserted into the retained distillation branch to enhance feature reuse with a small parameter increase. Batch normalization is removed to avoid disturbing low-level image statistics and to simplify inference. Experiments on DIV2K and benchmark datasets including Set5, Set14, and Urban100 show that PCF-IMDN reduces the parameter count from 715K to 430K and the computational cost from 158G to 92G compared with IMDN. The model achieves 4.75 ms latency on Set5 and 25.6 FPS for 1080P input, demonstrating its potential for real-time edge-oriented image enhancement.
This research introduces an enhanced version of the Cartographer algorithm for laser-based simultaneous localization and mapping (SLAM) to address common challenges in traditional LiDAR methods, such as incomplete point cloud features, low-quality data, and pose drift induced by measurement noise. The proposed approach incorporates an Adaptive Unscented Kalman Filter (AUKF) during sensor fusion, which effectively predicts and updates sensor measurements with adaptive noise optimization to mitigate interference during pose estimation. Additionally, in the point cloud processing stage, we improve upon standard voxel filtering by integrating weighted secondary screening inspired by particle filtering concepts, resulting in reduced redundancy and enhanced point cloud accuracy. Experimental evaluations were performed in both indoor and outdoor settings compared to the original Cartographer algorithm. Results from indoor tests showed a notable decrease of 17.2% in absolute translation error and 30.1% in absolute rotation error. Similarly, outdoor experiments demonstrated improvements of 25.8% and 28.9%respectively. These quantitative findings validate the effectiveness of the proposed algorithm in achieving significantly lower errors and superior mapping accuracy, showcasing its practical applicability.
The number of vehicles with network access keeps growing, and such growth brings more serious channel congestion and node overload problems to VANET systems, these problems will become much worse if the network only uses one cluster head node, to solve the above existing defects, this paper puts forward a dynamic clustering scheme named DCHVC, this scheme takes the motion similarity of vehicles as a reference standard, and uses a multi-index dual cluster head selection rule to raise the efficiency of vehicle data transmission, the whole scheme first divides stable vehicle clusters according to the similarity of vehicle speeds and the space distance between vehicles, the scheme also builds a network structure with two cluster heads, the main cluster node and the backup cluster node are selected through the CRITIC-TOPSIS multi-index evaluation method, this selection process will take both relative speed differences of vehicles and sustainable time of communication links into consideration, the system sets up a congestion control module based on cache queue occupancy, this module distributes data transmission tasks between the two cluster heads dynamically, meanwhile, adjustable cluster splitting and merging steps are used to keep the overall working state of clusters all the time, this paper builds a simulation environment by combining two simulation tools NS3 and SUMO, under test scenes with different vehicle moving speeds and different vehicle quantities, the operation performance of DCHVC is compared with three existing schemes including VWCA, KMRP and DCM, when the total number of vehicles in the test scene reaches 300, the DCHVC scheme can reach a network throughput of about 800 kbps, this value is 8.1% higher than the VWCA scheme, the average end-to-end transmission delay of this scheme can drop to around 0.69 seconds, which has a 37.3% delay reduction compared with VWCA, all data from simulation tests can prove that the DCHVC scheme can work steadily in V2X scenarios where vehicle positions change frequently, it can effectively improve the running stability of clusters and balance the data transmission load of every node.
Image style transfer, as a core cross-disciplinary technology in computer vision and non-photorealistic rendering, aims to preserve the semantic structure of the content image while transferring the artistic textures and brushstroke patterns of the style image. Early traditional methods struggled with complex scenes due to the semantic gap. With the advancement of deep learning, neural style transfer (NST) achieved a leap from "pixel-level statistics" to "feature-level perception." Crucially, while early optimization-based NSTs suffered from slow inference (less than 0.1 FPS), subsequent feed-forward networks enabled real-time processing (20-60 FPS). In recent years, Generative Adversarial Networks (GANs), Vision Transformers (ViTs), and denoising diffusion models have driven breakthroughs. Diffusion models, combined with parameter-efficient fine-tuning (e.g., LoRA) and distillation techniques, achieve sub-second high-fidelity generation with decoupled structural control. This paper systematically traces the technical evolution of image style transfer. The discussion unfolds across three dimensions: (1) establishing a classification system for datasets; (2) analyzing the intrinsic mechanisms of five major SOTA paradigms; and (3) comprehensively comparing mainstream evaluation metrics (such as LPIPS, FID, and Gram Matrix Distance) and model performance. Finally, this paper summarizes current challenges, such as high-resolution real-time inference, and projects future trends in video stylization and 3D scene transfer, providing a comprehensive roadmap for future researchers in this field.
Object detection in Unmanned Aerial Vehicle (UAV) scenarios faces significant challenges, including low resolution, a high prevalence of small objects, extreme scale variations, and dense occlusion. To address these issues, this paper proposes DART-DETR, a novel object detection architecture based on adaptive receptive fields and dynamic feature fusion. Specifically, this work introduces the DAFPN, which achieves content-adaptive fusion across multi-scale features via dual-path dynamic weight prediction, a cross-input redistribution mechanism, and lightweight context residual modeling. Furthermore, this model designs the ARMix module, which integrates learnable multi-kernel depthwise convolution with CGLU to enhance spatial-channel modeling capabilities. Extensive experiments on the VisDrone 2019 benchmark demonstrate that our DAFPN-RT-DETR achieves 50.0% mAP50 and 30.9% mAP50-95 with only 15.4M parameters and 59.7 GFLOPs. This work establishes a new state-of-the-art trade-off between efficiency and accuracy, providing an effective and generalizable solution for UAV-based small object detection.
Object detection in low-light environments poses a challenging task. Existing work involving visible and infrared light primarily focuses on fusion enhancements across multiple scales of the backbone network. This paper proposes and constructs the Prior Difference Interleaved Network (PDIN), a dual-stream detector based on the YOLOv11n framework. Its core innovation lies in the interleaved deep fusion strategy combining CPCA and SDI. Specifically, the model introduces a Channel-Prior Convolutional Attention (CPCA) module before fusion to pre-enhance and reduce redundancy in dual-modal features. Subsequently, a Semantic Difference Interaction (SDI) module is designed and proposed, whose core lies in converting semantic differences between modalities into dynamic weight signals that guide fusion, achieving difference-driven adaptive integration. By first optimizing feature quality and then performing dynamic difference-driven integration, PDIN significantly enhances model robustness. Extensive results on the VEDAI dataset demonstrate PDIN's effectiveness, ultimately improving the mAP50 performance metric from the baseline 47.2% to 52.3%. This study robustly validates the efficacy of explicitly leveraging modal differences and performing feature quality pre-enhancement in bimodal deep learning fusion.
To address the issues of low efficiency in manual modeling of virtual scenes and the high engineering threshold of existing AI generation technologies, an offline automatic virtual scene generation framework based on Generative Adversarial Networks (GAN) and the Unity engine is proposed. With the Pix2Pix conditional generation model as its core, this framework enables the automatic generation of indoor scene layouts. It completes the reconstruction from layouts to 3D scenes through the Unity parsing-generation module and adopts an "offline file collaboration" mode to reduce the cross-domain technical coupling. Experiments are conducted based on the SUNCG indoor dataset. The average Intersection over Union (IoU) between the layouts generated by the trained Pix2Pix model and the real layouts reaches 0.78. In a general PC environment, the Unity scene reconstruction module takes no more than 1.2 seconds to generate a 10×10m indoor scene, which improves the efficiency by over 99% compared with manual modeling. User experience tests show that this framework has a low operation threshold, and the generated scenes receive an average score of 4.1/5 in terms of layout rationality and practicality, which can meet the rapid development needs of scenarios such as game prototype design and educational virtual simulation.
Polar codes and their CRC-assisted derivatives have become important facilitators of ultra-reliable low-latency communication (URLLC) in 5G wireless networks, especially in the cases of applications where short packets have to be transmitted, and error rates must be nearly zero, such as autonomous vehicles, remote medical treatments, and automation of industrial devices. Although the dependability has been proved to be beneficial when the incorporation of codes of CRC-Polar is performed previously, the studies frequently analyze the qualities of the reliability and latency separately without taking into account the trade-offs inherent to 6G URLLC. This paper fills this gap by coming up with a single analytical model that explicitly characterizes the dependence of blocklength, their CRC length, and successive cancellation list size (size of successive cancellation) on the block error rate and the decoding latency. The it is a methodology that incorporates the finite-blocklength information theory, channel polarization principles and the analysis of instruments of CRC errors detection and is confirmed and validated over extensive simulation over realistic 6G parameters. The findings reveal that list decoding with the aid of CRC yields a significant reliability improvement in short packet cases, and the increases in decodinglatency with larger lists are nearly linear, although there is a large gain in reliability with a considerable like impact of CRC length. The paper defines limited optimum areas of the parameters under which URLLC requirements are met, the significance of co-designing of parameters. In general, the framework suggested will give viable considerations towards realizing hybrid codes of CRC-Polar mechanisms to reach ultra-high reliability without breaking the sub-milliseconds delay limitations in future 6G networks.
The escalating global mental health crisis has high- lighted a critical shortage of accessible therapeutic resources. While digital health interventions exist, many rely on static jour- naling or rudimentary rule-based chatbots that fail to capture the semantic nuance of complex human emotions. This paper presents the design, development, and evaluation of “Apricity,” a comprehensive AI-powered mental health companion application. Grounded in the principles of Cognitive Behavioral Therapy (CBT), the system provides users with an empathetic platform to track emotions and journal thoughts. The application is engineered as a scalable full-stack solution using the MERN stack (MongoDB, Express, React, Node.js) integrated with an asynchronous Python microservice for heavy inference tasks. A central contribution of this work is the implementation of a high-fidelity emotion recognition model. We fine-tuned the DeBERTa-v3 (Decoding-enhanced BERT with Disentangled Attention) architecture on the GoEmotions dataset, implementing a novel mapping strategy to aggregate 27 fine-grained labels into 5 core emotional categories (Joy, Sadness, Fear, Anger, Surprise). Experimental results demonstrate that our approach achieves a validation accuracy of 92.5% and a weighted F1-score of 0.83, significantly outperforming baseline models including BERT, RoBERTa, and traditional SVM classifiers. Furthermore, the system addresses the challenge of deploying large language models in consumer applications by utilizing a Job Queue architecture, ensuring real-time responsiveness.
To enhance the robustness of small object detection in UAV aerial imagery, this paper proposes an improved detection method based on the YOLOv11 architecture. Initially, to mitigate the occlusion of small-target features by surrounding background clutter, a Multi-scale Edge Information Enhancement module is introduced to amplify fine-grained local details. Subsequently, an efficient feature fusion architecture is constructed to achieve comprehensive integration of global contextual semantics with small-object feature representations. Furthermore, a Task-Dynamic Aligned Head (TDAH) based on shared convolutions is proposed to mitigate the inconsistency between the classification and regression tasks. Finally, a loss function named WSIoU, which incorporates dynamic focusing and shape-aware constraints, is introduced to reduce the interference of low-quality samples during model optimization. Experimental results on the VisDrone2019 dataset confirm that the proposed method not only achieves a 4.72% improvement in mAP@0.5 but also provides a practical and viable enhancement strategy for small object detection from a drone perspective.
