Fujian Polytechnic of Information Technology

Music emotion recognition and generation play a vital role in creating emotionally resonant experiences in human-computer interaction. Deep reinforcement learning (DRL) offers promising capabilities for dynamically modeling and generating music aligned with human emotional states. However, existing methods often struggle with rigid emotional mapping, a lack of real-time adaptability, and limited generalization across diverse musical contexts. To address these limitations, this study proposes an Emotion-Conditioned Deep Reinforcement Learning (EC-DRL) framework that integrates emotion-aware representations into the reward mechanism of a DRL agent. The system employs deep neural networks to extract high-level audio features and map them onto valence-arousal emotion space, guiding the music generation process through a reinforcement learning policy optimized for emotional congruence. This framework is applied in adaptive soundtrack generation for interactive applications such as video games, where music responds dynamically to emotional cues derived from gameplay and user interactions. Experimental results demonstrate that the EC-DRL framework significantly improves emotional accuracy, coherence, and user satisfaction compared to traditional sequence-based generation models. The proposed system shows the potential for creating emotionally intelligent music systems that can adapt in real-time and produce expressive musical output. EC-DRL achieves the mapping accuracy by 98%, emotional congruence score by 0.9%, real-time responsiveness by 280 ms, reward function optimization by 9.5%, audio feature extraction quality by 86%, policy convergence rate by 0.8%, user satisfaction score by 8.9%, and cross-domain generalization by 88%.

Chain-based routing protocols such as PEGASIS reduce redundant transmissions in Wireless Sensor Networks (WSNs), yet their static link structures and greedy neighbor selection often result in long communication links, uneven energy depletion, and premature network failure. Achieving both globally optimized chain construction and energy-balanced routing under diverse network deployments therefore remains a significant research challenge. To address these limitations, this paper proposes GWO-PEGASIS, a scalable two-stage optimization framework that integrates K-means clustering for spatial partitioning with the Grey Wolf Optimizer (GWO) to construct globally optimized intra- and inter-cluster chains under connectivity constraints. In the first stage, K-means forms spatially compact clusters based on node density, initial energy, and spatial distribution. In the second stage, GWO enables adaptive chainhead selection, balanced multihop forwarding, and improved spatial load distribution. Extensive simulations validate the effectiveness of the proposed method: GWO-PEGASIS shortens total chain length by 10.46% relative to PEGASIS, more than doubles the network lifetime, and further improves lifetime by 11.2% and 14.6% over EB-PEGASIS-SCL and MC-CRITIC-KM, respectively. It also achieves the lowest per-round energy consumption, the most balanced residual-energy distribution, and the highest throughput among all evaluated protocols. These results demonstrate that GWO-PEGASIS offers a robust, energy-efficient, and scalable routing solution for WSNs, highlighting the strong potential of swarm-intelligence-driven optimization for addressing complex topology design challenges in resource-constrained environments.

Human Pose Estimation (HPE) is a core component in real-time decision systems, supporting critical applications such as healthcare monitoring, autonomous driving, and sports analytics. While deep learning models—particularly CNNs and Transformer-based architectures—have significantly improved HPE accuracy, they remain vulnerable to adversarial perturbations that subtly distort keypoint localization, thereby undermining system reliability. To address this challenge, we propose regularization and pruning transferable adversarial attack (RPA), a novel framework designed to enhance the transferability of adversarial samples in Transformer-based HPE models. RPA integrates two synergistic strategies: gradient regularization, which suppresses dominant feature correlations to reduce overfitting, and adaptive weight pruning, which removes redundant parameters to reduce model-specific noise. This dual mechanism enables the generation of transferable adversarial attacks that are effective across diverse model architectures. Extensive experiments on state-of-the-art HPE networks demonstrate that RPA consistently outperforms existing attack methods. In white-box settings, RPA reduces average precision (AP) by 0.05-0.30; in black-box scenarios, it yields AP drops of 0.01-0.04. These findings expose critical vulnerabilities in IoT-enabled HPE applications and establish a new benchmark for evaluating adversarial robustness in real-time perception systems.

ABSTRACTBy integrating gene expression data, clinical features and multimodal data, we constructed a machine learning model capable of accurately predicting the prognosis of endometrial cancer patients. The study found that key genes related to lysosome‐dependent cell death exhibit significant expression pattern heterogeneity in endometrial cancer and are closely associated with immune cell infiltration and metabolic characteristics within the tumour microenvironment. Patients in the high‐risk group tend to have lower immune scores and a higher prevalence of immunosuppressive cell types, such as regulatory T cells and M2 macrophages, which may be linked to poorer prognosis and resistance to immunotherapy. Additionally, we discovered that the expression of lysosome‐dependent cell death‐related genes correlates with patients' sensitivity to chemotherapeutic drugs, providing new perspectives for personalised treatment of endometrial cancer. Through this study, we characterised the prognostic relevance of lysosome‐dependent cell death–related genes in endometrial cancer, and identified biomarkers with potential utility for risk assessment and therapeutic stratification.

The growing integration of AI technology in sports has transformed the way athletes train and monitor performance, yet traditional training systems often lack personalization and real-time adaptability. Common challenges include inaccurate performance tracking, limited feedback mechanisms, and inefficiency in processing large-scale physiological and motion data. To address these issues, this study proposes a Smart Sports Training Assistant System based on machine learning algorithms. In this context, this research proposes a novel Cognitive Monitoring System using the Sports training activities of the different individuals. The proposed framework consist of five components such as a) Intelligent Data Collection b) Data preparation for an effective training and testing the module c) Reservoir Computing Networks d) Model validation and Evaluation. Extensive experimentation is conducted using IoT (Internet of Things) based Sports activity benchmarks and performance metrics such as accuracy, precision, recall, F1-score are calculated. The empirical results were analysed and compared with the other existing state-of-the-art detection systems. The experimental outcomes demonstrated the proposed approach has achieved the 1.00 accuracy, precision of 0.99, recall of 0.99 and F1-score of 0.99 respectively The suggested research shows the promising direction for the involving this novel ensemble pervasive monitoring of the sports activities among the individuals.

Perovskite quantum dots (QDs) have attracted significant attention in the field of optoelectronics due to their potential in the area of nonlinear optics (NLO). However, these QDs are susceptible to aggregation or decomposition when exposed to moisture, light or elevated temperatures for extended periods. This inherent instability significantly restricts their practical applications. In this experiment, we synthesized composition-tunable CsPbX3 (X [Formula: see text] Cl, Br or I) QDs and encapsulated them in an acrylic resin matrix. This encapsulation provided robust protection, imparting high moisture resistance and thermal stability to the CsPbX3/Acrylic Resin composite. The composite retained strong photoluminescence and NLO properties even after seven days of storage in air or immersion in water, with structural integrity confirmed by transmission electron microscopy (TEM) and spectroscopic analysis. Quantitative optical limiting (OL) assessment under 532 nm pulsed laser revealed strong halide composition dependence: CsPbBr3/Acrylic Resin exhibited optimal performance, while incorporation of Cl[Formula: see text] or I[Formula: see text] reduced OL efficacy. The excellent NLO properties, combined with enhanced stability via acrylic resin encapsulation and tunability via halide composition control, indicate the composite’s strong potential for optical switching and limiting devices.

With the extensive growth of the web as well as cellular networks, secure multimedia transmission through cellular networks is needed. Currently, fifth-generation (5G) cellular networks are utilized to perform secure multimedia transmission. Numerous studies have been conducted to design efficient resource allocation approaches for secure video transmission in 5G cellular networks. However, this approach does not offer complete video security related to security against dynamic eavesdroppers or patent defilements. Thus, a resource allocation algorithm named Tasmanian devil whale optimization (TDWO) is introduced for secure video transmission in 5G networks. Here, the recorded educational videos are considered and are transmitted over 5G network transmission resources initially. The resources in 5G networks are allocated via the TDWO model by considering fitness parameters such as the data rate, achievable data rate, and quality of experience (QoE). Here, the deep convolutional neural network (DCNN) model is deployed for the prediction of the QoE in resource allocation. Moreover, extensive experiments are performed to identify the resource allocation performance of the designed TDWO model. The experimental results prove that the TDWO resource allocation algorithm yields significant experimental outcomes, with throughput, bit error rate (BER), QoE and fitness values of 25.557 Mbps, 0.021, 18.332 and 0.013, respectively.

The rapid expansion of edge computing has introduced significant challenges in trust management due to its decentralized and heterogeneous nature. Traditional trust models, such as rule-based or reputation-based systems, often fail to address the dynamic behavior of edge nodes and the constraints of computational resources. This study aims to develop a dynamic trust management model utilizing Q-learning, a reinforcement learning algorithm, to assess and adjust trust levels among edge nodes in real time. The model treats each node as an autonomous agent that learns from past interactions, refining trust decisions via a reward-based mechanism. Simulations were conducted in a dynamic mesh network environment using Python, with varying node behaviors such as honest, opportunistic, and malicious actions. The results show that the Q-learning-based model significantly outperforms traditional models, achieving trust classification accuracy of 91.2%, communication efficiency of 89.4%, and enhanced adaptability to behavioral changes. These findings suggest that reinforcement learning techniques, particularly Q-learning, can offer a more adaptive, lightweight, and resilient approach to trust management in decentralized edge computing systems, making them more robust to the inherent challenges of such environments. The proposed model has practical implications for improving security and efficiency in edge computing applications, supporting the development of more intelligent and reliable decentralized networks.

In recent years, the task of generating captions for videos has become a prominent research focus, with its main challenge being how to effectively capture essential semantic elements – such as objects, actions, and their spatial-temporal relationships – from abundant and redundant visual content. To address this challenge, earlier methods generally concentrate on either extracting representative clips across multiple frames (global level) or locating salient areas within single frames (local level). Many existing methods tend to ignore the fundamental hierarchical organization of videos, where identifying representative frames should come before locating informative regions. To tackle this limitation, we propose G2L, a hierarchical attention framework that (1) selects salient clips & frames via differentiable Gumbel Top-K sampling and (2) refines region-level context for caption generation. Extensive experiments conducted on the widely adopted benchmarks MSVD and MSR-VTT confirm that our method achieves notable improvements over existing state-of-the-art approaches. Ablations confirm that the global-to-local cascade and dual-branch optimization jointly account for the gain.