In the digital transformation of the judiciary, legal entity recognition is a foundational prerequisite for building intelligent judicial systems. To address the limited domain adaptability of generic pre-trained models and the computational burden of training large legal models, this paper proposes a lightweight yet effective legal entity recognition optimizer built upon the BERT-BiLSTM-CRF architecture. Empirical results demonstrate substantial gains in accuracy, efficiency, and deployability. With a legal-specialized adapter, the model attains 97.63% F1 on the CAIL2021-IE corpus, 2.68 pp above the baseline. Progressive unfreezing coupled with mixed-precision training substantially reduces training time and GPU memory footprint on a single consumer-grade GPU. Finally, the clause-aware attention mechanism further improves extraction quality on longer documents while reducing inference overhead in extended-context settings. Collectively, these innovations overcome the challenges of domain adaptation, resource overhead, and long-text processing in legal entity recognition, offering a practical deployment solution for resource-constrained deployments of legal AI.
To improve the efficiency and reliability of industrial digital instrument reading, this paper proposes an automatic recognition method based on YOLOv8. Aiming at the low efficiency and poor robustness of traditional manual and rule-based methods, YOLOv8 is used to accurately detect the digital display area of instruments, benefiting from its fast inference speed and strong adaptability to complex environments. Subsequently, image preprocessing operations, including grayscale conversion, denoising, binarization, and morphological processing, are applied to enhance digital features, and individual digits are segmented and recognized using a ResNet34 classifier. Experimental results show that the proposed method achieves a detection accuracy of over 98.20% for digital display areas and a digit recognition accuracy of over 98.60%, demonstrating good robustness and practical applicability in complex industrial scenarios.
Spam email has long threatened communication security and work efficiency. To address this, we design and implement an email spam-detection agent that integrates the DeepSeek large language model, the Dify agent-development platform, and a fine-tuned BERT model. The system uses BERT as the core classifier, leveraging its strengths in semantic understanding and deep feature extraction; it adopts a binary scheme (label 0 = ham, 1 = spam), and fine-tuning enables effective recognition of email text features. Meanwhile, the DeepSeek LLM is introduced to exploit its capabilities in reasoning and generation: for ham messages, DeepSeek produces key-point summaries, and for spam it provides risk explanations and safety recommendations. With Dify’s tool orchestration and human–computer interaction interface, the system automates the entire pipeline from parsing email content to intelligent decision-making and interactive feedback, forming an end-to-end agentic framework for spam detection.For experiments, we train and validate on the Kaggle email dataset (33,715 messages: 17,170 spam / 16,545 ham), using a 70%/15%/15% train/validation/test split. On the spam-detection task, the system achieves 99.49% accuracy; precision, recall, and F1 reach 99.42%, 99.57%, and 99.49%, respectively. These results demonstrate excellent detection performance and strong generalization. In summary, the proposed DeepSeek–Dify–BERT integrated agent effectively safeguards user communications, reduces potential information-security risks, and substantially improves the intelligence and automation of the detection workflow.
The performance of visual SLAM is strongly influenced by the quality of front-end feature detection and correspondence matching. To improve ORB-SLAM3 in weak-texture environments, under feature clustering, and in the presence of mismatches, this paper optimizes the front-end pipeline in three stages. First, an adaptive threshold is introduced into FAST detection to improve keypoint extraction in weak-texture regions. Second, an improved quadtree-based distribution strategy is adopted to reduce feature over-concentration in strongly textured areas while retaining more valid keypoints in weak-texture regions. Third, PROSAC is used for correspondence verification to remove mismatches with lower iterative cost than conventional random sampling. The improved front-end is integrated into ORB-SLAM3 and evaluated on the EuRoC Vicon Room1 03 sequence. Experimental results show clear gains in feature extraction and matching quality, reducing the mean Absolute Trajectory Error (ATE) by 81.8% and the mean Relative Pose Error (RPE) by 89.4% relative to the baseline, thereby improving localization and mapping accuracy.
This paper proposes an AI-enhanced QKD protocol, which uses machine learning-based adaptive control to dynamically optimize error correction and entanglement quality in view of the dynamics of network conditions. The proposed framework integrates the AI prediction model with Low-Density Parity-Check codes and entanglement swapping, aiming at the intelligent regulation of photon loss, QBER, and the key generation rate. An AI model predicts real-time channel noise and thus dynamically adjusts LDPC parameters and entanglement fidelity thresholds to reach a self-optimizing QKD system. Simulation results show that the proposed protocol has improved the achievable transmission distance by up to 120% (up to 220 km) and suppressed QBER by 65% compared to standard BB84 and previously optimized QKD protocols. This work underscore the crucial steps toward a more autonomous, scalable, and resilient quantum network, enabling secure communication over global distances.
To address the problem of excessive outlier errors caused by noise in indoor non-line-of-sight (NLOS) environments—particularly in positioning and robot localization—two Chan–Taylor cooperative algorithms are proposed and implemented to suppress NLOS-induced errors. The first approach integrates Kalman filtering for error mitigation, while the second reconstructs distance measurements based on statistical characteristics. By combining multiple positioning algorithms, the proposed methods effectively reduce the impact of high noise levels and NLOS interference. Dynamic and static positioning tests were conducted to evaluate the accuracy of both schemes. The results indicate that localization errors in NLOS environments can be significantly reduced by both approaches, with the NLOS error suppression algorithm exhibiting superior performance and adaptability.
In response to the limitations of traditional obstacle avoidance algorithms for unmanned vehicles in dealing with unknown obstacles and complex dynamic environments, this paper proposes Q Mixing Network and Video Delivery Network reinforcement learning algorithms, specifically for the research of obstacle avoidance decision-making for unmanned vehicles. By constructing a mapping relationship between local function values and global function values, it is possible to guide obstacle avoidance decisions for unmanned vehicles based on the decomposed function values. Experimental verification was conducted on a simulation platform based on ROS+Gazebo, and compared with the Quantile Regression for Reinforcement Learning algorithm in the same testing environment. The results showed that QMIX and VDN algorithms were more adaptable to complex map environments during training, with obstacle avoidance success rates increased by 16.9% and 18.1%, respectively, effectively improving the obstacle perception and avoidance capabilities of unmanned vehicles.
To address the challenges in efficiently modeling and simulating Electromagnetic Compatibility (EMC) for complex electrical and electronic systems, a system-level EMC analysis method based on cascaded multi-port network theory is proposed. According to the topological structure of the system, this method decomposes the complex system into multiple cascaded multi-port network modules, where electromagnetic energy transmission and coupling occur via ports. In this method, the Electromagnetic Interference (EMI) transfer characteristics of each module are described by the impedance matrix of its internal circuitry. The voltage and current coupling relationships at system ports are calculated using the cascaded impedance matrix of the multi-stage network combined with specific boundary conditions, thereby significantly simplifying the modeling and simulation process. Numerical validation performed on a three-stage cascaded system across the 10–100 MHz frequency band demonstrates that the proposed method achieves high fidelity. Compared with full-wave simulations, the maximum absolute error for the current response is limited to 0.7 mA (occurring at 10 MHz), and for the voltage response is merely 0.037 V, with the average current error maintained at approximately 0.1 mA. The relative error for both responses at critical ports consistently remains within 0.2%. Furthermore, both modeling complexity and calculation time are substantially reduced, providing a feasible technical approach for the EMC modeling and simulation of complex systems.
To address the issues of uneven texture distribution, background mis-migration, and training imbalance in traditional CycleGAN for style transfer of non-paired horse and zebra images, this study proposes an improved model integrating dynamic attention mechanisms, semantic segmentation constraints, and adaptive training strategies. By embedding lightweight space-channel hybrid attention modules in generator residual blocks, the model enhances feature extraction in target regions. A lightweight semantic segmentation network is introduced to enforce local style transfer constraints, preventing redundant background migration. A two-stage adaptive loss weight adjustment strategy is designed to improve training stability. Experiments on the horse2zebra dataset using the PyTorch platform demonstrate that the improved model generates zebra stripes that conform to the subject structure, with non-target region pixel changes below 15.8%. Compared to the original CycleGAN, the model achieves an 8.3% SSIM improvement and 1.5dB PSNR enhancement. This model effectively resolves core limitations of traditional methods, providing a superior solution for style transfer of non-paired animal images.
Gamma-ray burst (GRB) detection is crucial for triggering rapid follow-up observations and enabling subsequent multi-wavelength studies, yet traditional methods are limited by their difficulty in modeling long-range temporal dependencies in noisy, non-stationary data. In this work, we propose GRBNet, a self-attention–based neural network for GRB detection from same-source multi-detector time-tagged event (TTE) sequences. GRBNet leverages multi-head self-attention to adaptively aggregate discriminative evidence over the full observation window, explicitly capturing long-term dependencies and multi-pulse structures, and integrates complementary responses from multiple detectors under a unified tokenization scheme to improve robustness against background drift and low SNR in individual instruments. Experiments on real Fermi/GBM observations show strong detection performance (recall=1.00, precision=1.00) under a stringent triggering threshold 0.99, indicating reliable performance for weak and morphologically complex GRB events.
