A GPU resource prediction method based on the SGRU-AATransformer hybrid modelHan, Boxiao; Ni, Weichen; Xu, Fei; Shen, Lei
doi: 10.1117/12.3068791pmid: N/A
With the extensive application of GPUs in artificial intelligence, high-performance computing, and other fields, especially their critical role in future cloud-based wireless networks, accurately predicting GPU load has become crucial. This paper proposes a novel hybrid model—SGRU-AATransformer, aiming to improve the accuracy of GPU load prediction. The model combines the Shared Gated Recurrent Unit (SGRU) for capturing short-term nonlinear dynamic changes and the Approximate Attention Transformer (AATransformer) for modeling long-term dependencies. Experimental results demonstrate that SGRU-AATransformer performs well in both short-term and long-term predictions, significantly outperforming existing single and hybrid prediction models.
Research on storage optimization method for automated stereoscopic warehouse based on genetic algorithmChai, Yajie; Hou, Jinbao; Wu, Tao
doi: 10.1117/12.3068818pmid: N/A
The deepening application and development of automated stereoscopic warehouses not only improve the efficiency of goods entering and exiting the warehouse, but also face the problem of dynamic storage space allocation optimization, which not only directly affects the average travel time of stacker cranes, but also relates to the overall operational efficiency of stereoscopic warehouses. In response to this, this article combines the main problems of automated stereoscopic warehouses, constructs a storage space allocation optimization problem model, and improves the traditional genetic algorithm to complete the solution of the model.
Minimizing end-to-end delay in three-layer MEC with dual-input requirementFan, Yunfei; Gao, Yang; Lian, Qingqun; Wang, Meiling
doi: 10.1117/12.3068439pmid: N/A
Most mobile edge computing (MEC) scenarios assume that all the input data of offloading tasks is transmitted from users. However, we notice that a more general situation, i.e., at least a part of input data, as common knowledge, should be provided from outside environments (e.g., the Internet). In this paper, we formulate the end-to-end delay in the High Altitude Platform Station (HAPS) and Roadside unit (RSU) assisted MEC, where the input data of these tasks are provided from both HAPS and users/nodes. Then we adopt the branch and bound method to relax this mixed-integer nonlinear programming (MINLP) problem. The Karush-Kuhn-Tucker (KKT) conditions and the interior-point method are utilized to analyze and optimize the relaxed problem. Experimental results show that our scheme can significantly reduce the end-to-end delay of data transmission and task offloading compared to other benchmark schemes.
A data filling method for bank direct selling missing data based on an improved EM algorithmLi, Xiaoyan; Xu, Yong
doi: 10.1117/12.3068649pmid: N/A
Aiming at the problem of missing data in bank marketing, this paper proposes an alternative approach that involves combining the EM algorithm with Gibbs sampling. Given a set of incomplete data, the proposed approach uses samples obtained by Gibbs sampling to replace missing values in the expectation step of the EM algorithm. Afterwards, the likelihood function is used to complete the expectation maximization iterative update. The missing data is filled in by repeated iterations to obtain the maximized estimate as the reconstructed value. Simulation experiments using a bank marketing campaign dataset shows that the proposed algorithm perform better than the EM algorithm in data filling while effectively addressing the susceptibility of EM to initial value settings.
Safety helmet wearing detection algorithm in complex scenarios based on improved YOLOv8nSheng, Peng; Zhang, Min; Zhu, Zixuan; Jin, Congqian; Fang, Jiwei; Jiang, Wenhao
doi: 10.1117/12.3068404pmid: N/A
To address the challenges faced by safety helmet detection in complex environments, particularly the impact of fog, dust, occlusion, and varying lighting conditions , existing algorithms exhibit poor performance. To overcome these issues, this paper proposes an improved YOLOv8n algorithm. First, the Channel-Grouped Multi-Scale Convolution Module (CMSC) is integrated into C2f, using multi-scale convolution to enhance feature representation in foggy and dusty environments, reducing blurring. Second, the detection head is improved by switching from parallel to serial structure and introducing Partial Convolution (PConv), enhancing occlusion handling. Finally, WIoU v3 optimizes bounding box regression under varying lighting conditions via adaptive weight adjustment. Compared to the baseline, the method improves accuracy by 0.3%, recall by 2.7%, [email protected] by 2.2%, while reducing parameters to 2.5M and GFLOPs by 2.7G, achieving a balance between efficiency and accuracy.
A threshold multiserver authentication protocol based on multifactorQin, Shihan; Zhang, Rui; Xiao, Yuting
doi: 10.1117/12.3068544pmid: N/A
With the development of cloud computing, users frequently access various cloud-based services. Single sign-on (SSO) mechanisms, which delegate authentication to multiple servers using a only password and issue tokens for subsequent access, have become widely adopted. However, password-based SSO protocols are vulnerable to offline guessing attacks and suffer from single points of failure due to reliance on a single authentication server. To address these issues and enhance authentication flexibility, we propose a threshold multi-server authentication protocol based on multi-factor. The proposed protocol: (1) enables flexible user authentication with any 𝑡 out of 𝑛 registered factors; (2) extends traditional single-server authentication to a multi-server setting, effectively mitigating single points of failure and user factor leakage risks. Security analysis and performance evaluation further demonstrate the protocol’s security, robustness, and efficiency. This makes it particularly suitable for practical applications such as cloud services, enterprise authentication, and IoT environments, where both security and flexibility are essential.
A study on 3D printer fault diagnosis based on the MLP-XGB hybrid modelFang, Lin
doi: 10.1117/12.3068467pmid: N/A
Traditional methods often rely on manual feature extraction and expert knowledge, which are labor-intensive and lack scalability. This study proposes a novel hybrid model based on the combination of a Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (XGB) for diagnosing faults in 3D printers. The proposed MLP-XGB model leverages the nonlinear modeling capabilities of MLP and the efficient feature selection of XGB to enhance diagnostic accuracy. Experimental results show that the MLP-XGB model achieved high accuracy rates of 97.447%, 97.569%, and 97.503% on the training, validation, and test sets, respectively, with an F1 score close to 0.975. The 5-fold cross-validation yielded a mean accuracy of 96.3% and a standard deviation of 0.0094, demonstrating strong stability and generalization capability. The AUC values were close to 1, indicating excellent discrimination between positive and negative samples. Comparisons with other advanced models, such as iTransform-LSTM-GXB and CNN-GRU-GXB, show that the proposed model has potential. Future work will focus on data augmentation, model fusion, innovative deep learning architectures, model interpretability, and real-time fault diagnosis to further enhance the performance and practicality of 3D printer fault diagnosis.
Supervised contrastive probability learning and sharpness-aware minimization model for multimodal emotion recognition in conversationsLiu, Hailong; Rao, Wenbi
doi: 10.1117/12.3068505pmid: N/A
Multimodal Emotion Recognition in Conversation (MERC) aims to extract emotional information from conversations, but its performance is limited by the quality of the dataset. In real life, data often exhibits data imbalance problem. Current solutions to this problem, such as data resampling and category-balanced loss functions, have their limitations. This paper proposes a model named SCPL-SAM, which combines supervised contrastive probability learning and sharpness-aware minimization techniques to effectively address the data imbalance issue. By modeling the sample feature distribution with Angular Central Gaussian distribution for data augmentation, the model reduces the number of samples and computational resources required for contrastive learning, and encourages the model to escape “saddle points,” thereby enhancing classification performance. On the MELD and IEMOCAP datasets, the SCPL-SAM model outperforms other models, demonstrating its effectiveness in dealing with the data imbalance issue.
Hardware implementation of real-time NLM algorithmFan, Gongyu; Zhou, Wenbiao; Lu, Fengchi; Zhu, Boyu; Zhou, Yongxiang
doi: 10.1117/12.3068598pmid: N/A
Recent years, Image Signal Processor (ISP) has been wildly used in our daily lives. Non-local means (NLM) is one of the most robust noise reduction algorithms, but its computational complexity makes it difficult to be used for the realtime processing. In this paper, we propose a hardware implementation of real-time NLM algorithm and discuss some details of the implementation. Eventually we synthesized and implemented it on an FPGA to analyze the resource utilization and the performance and draw some common conclusions.
Self-supervised learning based on local attention encoding module for human activity recognition with wearable dataChu, Jianping; Zhang, Yanmei; Wang, Xingbo; Chen, Wenchen
doi: 10.1117/12.3068426pmid: N/A
A self-supervised transfer learning method based on a local attention encoding module (LAEM) is proposed for human activity recognition using wearable devices. This method effectively captures spatiotemporal features through the local attention mechanism and leverages a self-supervised learning strategy to extract general features from unlabeled data, significantly reducing reliance on labeled data. Experimental results indicate that the proposed method achieves an average F1 score improvement of 2.7% across multiple target datasets, with the maximum improvement reaching 4.3%. By thoroughly fine-tuning the model structure, the method further enhances the accuracy and transferability of activity recognition, demonstrating outstanding performance in cross-dataset transfer learning and small dataset scenarios. Additionally, the approach optimizes feature representation for target tasks and validates its adaptability and generalization capabilities under data-scarce conditions