作者：
Fan Yang;Tan Zhu*;Jing Huang;Zhilin Huang;Guoqi Xie
期刊：
COMPUTER SPEECH AND LANGUAGE,2026年95:101818 ISSN：0885-2308
通讯作者：
Tan Zhu
作者机构：
[Fan Yang; Tan Zhu; Zhilin Huang] College of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha, Hunan, 41004, PR China;[Jing Huang] School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, Hunan, 411201, PR China;[Guoqi Xie] Key Laboratory for Embedded and Network Computing of Hunan Province, College of Computer Science and Electronic Engineering, Hunan University, Changsha, Hunan 410082, PR China
通讯机构：
[Tan Zhu] C;College of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha, Hunan, 41004, PR China
摘要：
Text classification is an important topic in natural language processing. In recent years, both graph kernel methods and deep learning methods have been widely employed in text classification tasks. However, previous graph kernel algorithms focused too much on the graph structure itself, such as the shortest path subgraph,while focusing limited attention to the information of the text itself. Previous deep learning methods have often resulted in substantial utilization of computational resources. Therefore,we propose a new graph kernel algorithm to address the disadvantages. First,we extract the textual information of the document using the term weighting scheme. Second,we collect the structural information on the document graph. Third, graph kernel is used for similarity measurement for text classification. We compared eight baseline methods on three experimental datasets, including traditional deep learning methods and graph-based classification methods, and tested our algorithm on multiple indicators. The experimental results demonstrate that our algorithm outperforms other baseline methods in terms of accuracy. Furthermore, it achieves a minimum reduction of 69% in memory consumption and a minimum decrease of 23% in runtime. Furthermore, as we decrease the percentage of training data, our algorithm continues to achieve superior results compared to other deep learning methods. The excellent experimental results show that our algorithm can improve the efficiency of text classification tasks and reduce the occupation of computer resources under the premise of ensuring high accuracy.
Text classification is an important topic in natural language processing. In recent years, both graph kernel methods and deep learning methods have been widely employed in text classification tasks. However, previous graph kernel algorithms focused too much on the graph structure itself, such as the shortest path subgraph,while focusing limited attention to the information of the text itself. Previous deep learning methods have often resulted in substantial utilization of computational resources. Therefore,we propose a new graph kernel algorithm to address the disadvantages. First,we extract the textual information of the document using the term weighting scheme. Second,we collect the structural information on the document graph. Third, graph kernel is used for similarity measurement for text classification.
We compared eight baseline methods on three experimental datasets, including traditional deep learning methods and graph-based classification methods, and tested our algorithm on multiple indicators. The experimental results demonstrate that our algorithm outperforms other baseline methods in terms of accuracy. Furthermore, it achieves a minimum reduction of 69% in memory consumption and a minimum decrease of 23% in runtime. Furthermore, as we decrease the percentage of training data, our algorithm continues to achieve superior results compared to other deep learning methods. The excellent experimental results show that our algorithm can improve the efficiency of text classification tasks and reduce the occupation of computer resources under the premise of ensuring high accuracy.
期刊：
International Journal of Autonomous and Adaptive Communications Systems,2025年18(1):45-66 ISSN：1754-8632
作者机构：
[Zhilin Chen; Jiaohua Qin] College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, 410004, China
关键词：
RDH;reversible data hiding;prediction error;Huffman coding;encrypted images;extended run-length coding
摘要：
To reduce prediction errors and create more room for embedding data, the paper proposes a reversible data hiding (RDH) in encrypted images scheme based on histogram shifting and prediction error block coding. Firstly, the histogram of the prediction error image is shifted according to the signs of prediction errors. Next, the prediction error plane is partitioned into uniformly sized blocks, and these blocks are labelled as three types: an all-zero block, a block containing only one 1, and a block containing more than one 1. These three types of blocks are compressed using labelling, binary tree coding, and Huffman coding, respectively. To better compress the label map, an improved extended run-length coding is proposed. Finally, the image is secured by encryption and the secret data is hidden within it. The experimental results indicate a significant improvement in the embedding rate of the scheme compared to other schemes.
摘要：
Cloud environments enhance diffusion model efficiency but introduce privacy risks, including intellectual property theft and data breaches. As AI-generated images gain recognition as copyright-protected works, ensuring their security and intellectual property protection in cloud environments has become a pressing challenge. This paper addresses privacy protection in diffusion model inference under cloud environments, identifying two key characteristics-denoising-encryption antagonism and stepwise generative nature-that create challenges such as incompatibility with traditional encryption, incomplete input parameter representation, and inseparability of the generative process. We propose PPIDM (<bold>P</bold>rivacy-<bold>P</bold>reserving <bold>I</bold>nference for <bold>D</bold>iffusion <bold>M</bold>odels), a framework that balances efficiency and privacy by retaining lightweight text encoding and image decoding on the client while offloading computationally intensive U-Net layers to multiple non-colluding cloud servers. Client-side aggregation reduces computational overhead and enhances security. Experiments show PPIDM offloads 67% of Stable Diffusion computations to the cloud, reduces image leakage by 75%, and maintains high output quality (PSNR = 36.9, FID = 4.56), comparable to standard outputs. PPIDM offers a secure and efficient solution for cloud-based diffusion model inference.
通讯机构：
[Li, T ] H;Hunan Univ Sci & Technol, Sch Comp Sci & Engn, Xiangtan 411201, Hunan, Peoples R China.
关键词：
Short texts typically refer to texts of relatively short length;generally not exceeding 160 characters. Short text classification (Chakraborty & Singh;2022) is a form of text mining technology and a significant research direction in natural language processing (NLP) (Hirschberg & Manning;2015). With the rapid development of the internet;a large volume of information in the form of short texts such as news tags;reviews;instant messages;and tweets has emerged on online platforms. These short texts contain a wealth of helpful information that contributes to establishing services in various fields;such as news classification;social media (Kateb & Kalita;sentiment analysis (Balomenos et al.;e-commerce;and spam filtering. As a result;with the growing depth of related research;short text classification has emerged as a critical challenge that urgently needs to be addressed in information retrieval. However;short text classification faces several challenges that complicate the task. First;short texts tend to be brief;with sparse semantic features and weak contextual connections;often requiring additional feature information to assist in classification (Wu;2023). For example;Hua et al. (2024) integrated text;entities;and words into a heterogeneous graph convolutional network;representing features through a word graph and enhancing word features with BiLSTM;thereby enabling document category prediction. However;due to length limitations;GNNs typically do not incorporate additional information when processing short texts;treating each short text as a single node in the graph. This leads to insufficient feature information;which in turn affects performance. Therefore;enhancing feature information and semantic relevance becomes crucial. Second;in rapidly changing text data environments (such as social media and news);expert knowledge required to organize and label training documents properly is often limited;expensive;and sometimes even unavailable. In some cases;training text classifiers using well-organized and accurately labeled documents is prohibitively costly;making unsupervised learning a crucial solution. Unsupervised tasks allow for fast classification without labeled data;avoiding the time-consuming and error-prone process of manual annotation. For example;Cameron-Steinke (2019) employed a novel text expansion approach in an unsupervised manner;whereby new words are generated directly for each input sentence. The newly generated keywords are formed within the hidden states of a pre-trained language model;which are then used to create expanded pseudo-headlines. However;it has limitations in terms of the accuracy and redundancy of the expanded content. In addition;Mishra (2024) compared and analyzed unsupervised text classification techniques;especially Latent Dirichlet Allocation (LDA) and BERTopic;and applied these topic modeling techniques to very short documents. However;since their evaluation datasets are based on specific scenarios;the generalization of these methods is weak. Therefore;how to effectively classify large amounts of unlabeled data using unsupervised methods is a research-worthy problem. To address the aforementioned issues;we propose a novel approach called UEDE: The Unsupervised Short Text Classification Method based on GCN Encoder-Decoder and Local Enhancement. This method uses unsupervised learning to cleverly integrate the GCN encoder–decoder and local enhancement technology under the autoencoder framework to achieve accurate short-text classification. Specifically;we first design the MRFasGCN modeling approach within the autoencoder framework;learning the hierarchical representations of node attributes and the network’s deep structure. We model and infer the relationships between nodes by integrating these deep representations with network structural information;thereby capturing text features and semantic information more effectively. Then;we introduce a dual decoder suited for unsupervised learning;which alleviates the issue of short text feature sparsity from multiple perspectives by reconstructing the topology and node attributes. Furthermore;we further enhance the latent semantic module locally by constructing an information fusion graph;ensuring that each pair of connected nodes shares a similar module distribution and complementary features. This improves the model’s ability to understand the local context of the text while addressing the overfitting issues caused by short text feature sparsity;significantly boosting the model’s performance and classification accuracy. The main contributions of this article can be summarized as follows: • We propose an unsupervised learning-based short text classification method;utilizing multiple convolutional layers as encoders to learn short text attribute networks’ deep representations and structural information. This approach facilitates module partitioning and inference;more effectively capturing text features and semantic information. • We design a dual decoder structure that;through unsupervised learning;separately reconstructs the topology and node attributes while accounting for the heterogeneity of node degrees. This approach compensates for insufficient short text feature information from multiple perspectives. • We introduce a local enhancement method that enables nodes to share more common neighbors and similar node attributes;strengthening the connections between related nodes. This improves the model’s understanding of the text’s local context while alleviating the overfitting issues caused by short text feature sparsity. • We perform comprehensive experiments on real-world datasets encompassing news articles;concise comments;and search snippets to assess the efficacy of our model in comparison to fourteen baseline approaches. The experimental findings unequivocally establish that our model surpasses the current state-of-the-art baseline methods on the benchmark datasets. We propose an unsupervised learning-based short text classification method;utilizing multiple convolutional layers as encoders to learn short text attribute networks’ deep representations and structural information. This approach facilitates module partitioning and inference;more effectively capturing text features and semantic information. We design a dual decoder structure that;through unsupervised learning;separately reconstructs the topology and node attributes while accounting for the heterogeneity of node degrees. This approach compensates for insufficient short text feature information from multiple perspectives. We introduce a local enhancement method that enables nodes to share more common neighbors and similar node attributes;strengthening the connections between related nodes. This improves the model’s understanding of the text’s local context while alleviating the overfitting issues caused by short text feature sparsity. We perform comprehensive experiments on real-world datasets encompassing news articles;concise comments;and search snippets to assess the efficacy of our model in comparison to fourteen baseline approaches. The experimental findings unequivocally establish that our model surpasses the current state-of-the-art baseline methods on the benchmark datasets. The remainder of the paper is structured as follows: Section 2 reviews previous work. Section 3 describes our proposed method and model;including the unsupervised short text classification model and local enhancement method. In Section 4;we perform comprehensive experiments on the datasets and analyze the outcomes. Lastly;Section 5 concludes the paper;offering insights into the future research directions.
摘要：
Like all fields of data science, short text classification seeks to achieve high-quality results with limited data. Although supervised learning methods have made notable progress in this area, they require much-labeled data to achieve adequate accuracy. However, in many practical applications, labeled data is scarce, and manual labeling is not only time-consuming and labor-intensive but also expensive and may require specialized expertise. Therefore, this paper addresses the challenge of insufficient labeled data through unsupervised methods while ensuring the effective extraction of semantic features from the text. Building on this objective, we propose a novel unsupervised short text classification method within the framework of autoencoders . Specifically, we first design the MRFasGCN encoder and derive the relationships between nodes in its hidden layers, thereby enhancing the capture of text features and semantic information. Furthermore, we construct a dual-node-based decoder that reconstructs the topology and node attributes unsupervised. This approach compensates for feature deficiencies from multiple perspectives, alleviating the issue of insufficient features in short texts. Finally, we propose a localized enhancement method that integrates node features and topology, strengthening the connections between relevant nodes. This improves the model’s understanding of the text’s local context while mitigating the overfitting issues caused by feature sparsity in short texts. Extensive experimental results demonstrate the pronounced superiority of our proposed UEDE model over existing methods on the dataset, validating its effectiveness in short-text classification. Our code is submitted in https://github.com/w123yy/UEDE .
Like all fields of data science, short text classification seeks to achieve high-quality results with limited data. Although supervised learning methods have made notable progress in this area, they require much-labeled data to achieve adequate accuracy. However, in many practical applications, labeled data is scarce, and manual labeling is not only time-consuming and labor-intensive but also expensive and may require specialized expertise. Therefore, this paper addresses the challenge of insufficient labeled data through unsupervised methods while ensuring the effective extraction of semantic features from the text. Building on this objective, we propose a novel unsupervised short text classification method within the framework of autoencoders . Specifically, we first design the MRFasGCN encoder and derive the relationships between nodes in its hidden layers, thereby enhancing the capture of text features and semantic information. Furthermore, we construct a dual-node-based decoder that reconstructs the topology and node attributes unsupervised. This approach compensates for feature deficiencies from multiple perspectives, alleviating the issue of insufficient features in short texts. Finally, we propose a localized enhancement method that integrates node features and topology, strengthening the connections between relevant nodes. This improves the model’s understanding of the text’s local context while mitigating the overfitting issues caused by feature sparsity in short texts. Extensive experimental results demonstrate the pronounced superiority of our proposed UEDE model over existing methods on the dataset, validating its effectiveness in short-text classification. Our code is submitted in https://github.com/w123yy/UEDE .
作者机构：
[Long, Wei; Wang, Kailiang] State Key Laboratory of Tree Genetics and Breeding, Key Laboratory of Tree Breeding of Zhejiang Province, Research Institute of Subtropical Forestry, Chinese Academy of Forestry, Hangzhou 311400, Zhejiang, China;[Yang, Fan; Du, Jiayi] College of Computer and Information Engineering, Central South University of Forestry & Technology, Changsha 410004, Hunan, China;[Yu, Chunlian] Changshan Country Oil Tea Industry Development Center, Changshan 324299, China;[Lyu, Leyan] College of Hydraulic Engineering, Zhejiang Tongji Vocational College of Science and Technology, Hangzhou 311231, Zhejiang, China;[Dong, Zhipeng] State Key Laboratory of Tree Genetics and Breeding, Key Laboratory of Tree Breeding of Zhejiang Province, Research Institute of Subtropical Forestry, Chinese Academy of Forestry, Hangzhou 311400, Zhejiang, China<&wdkj&>College of Computer and Information Engineering, Central South University of Forestry & Technology, Changsha 410004, Hunan, China
通讯机构：
[Kailiang Wang] S;State Key Laboratory of Tree Genetics and Breeding, Key Laboratory of Tree Breeding of Zhejiang Province, Research Institute of Subtropical Forestry, Chinese Academy of Forestry, Hangzhou 311400, Zhejiang, China
摘要：
Camellia oleifera , an essential woody oil tree species, produces camellia oil, which is highly nutritious and edible after seed pressing. The demand for camellia oil has increased in recent years, so accurate estimates of camellia oil production are key to fruit harvesting and price setting. The rapid development of computer vision and deep learning techniques has found wide applications in agriculture, demonstrating significant fruit detection capabilities. This study proposes a surface fruit detection algorithm in C.oleifera based on the improved YOLOv8 model, with lightweight RepViT as the backbone of the model, a small object detection layer with p2 layer added on the multi-scale basis of FPN and PAN, and a network structure optimized by Shape IoU loss function. Meanwhile, the improved ByteTrack algorithm based on ByteTrack is utilized for multi-scale small-target tracking. Using LightGBM and linear regression to establish the relationship between the number of surface fruits and the number of individual fruits, combined with the number of typical sample trees and the average single fruit quality in the dataset, a production estimation model for the automatic detection of C.oleifera is constructed. The mAP value of the improved YOLOv8 algorithm on the C.oleifera fruit dataset reached 86.21 %, an improvement of 4.51 % compared to the original model. Combined with average single-fruit mass and the number of representative sample trees, these counts yielded whole-tree production estimates.The approach produced a fruit-number prediction R ²of 0.945 and a yield- estimation R ²of 0.902 across our sample trees, demonstrating strong linear relationships and validating the pipeline for automated C. oleifera yield forecasting.
Camellia oleifera , an essential woody oil tree species, produces camellia oil, which is highly nutritious and edible after seed pressing. The demand for camellia oil has increased in recent years, so accurate estimates of camellia oil production are key to fruit harvesting and price setting. The rapid development of computer vision and deep learning techniques has found wide applications in agriculture, demonstrating significant fruit detection capabilities. This study proposes a surface fruit detection algorithm in C.oleifera based on the improved YOLOv8 model, with lightweight RepViT as the backbone of the model, a small object detection layer with p2 layer added on the multi-scale basis of FPN and PAN, and a network structure optimized by Shape IoU loss function. Meanwhile, the improved ByteTrack algorithm based on ByteTrack is utilized for multi-scale small-target tracking. Using LightGBM and linear regression to establish the relationship between the number of surface fruits and the number of individual fruits, combined with the number of typical sample trees and the average single fruit quality in the dataset, a production estimation model for the automatic detection of C.oleifera is constructed. The mAP value of the improved YOLOv8 algorithm on the C.oleifera fruit dataset reached 86.21 %, an improvement of 4.51 % compared to the original model. Combined with average single-fruit mass and the number of representative sample trees, these counts yielded whole-tree production estimates.The approach produced a fruit-number prediction R ²of 0.945 and a yield- estimation R ²of 0.902 across our sample trees, demonstrating strong linear relationships and validating the pipeline for automated C. oleifera yield forecasting.
会议论文集名称：
Communications in Computer and Information Science
关键词：
Image manipulation detection;frequency information
摘要：
The goal of Image Manipulation Detection (IMD) is to identify and locate manipulated regions within images. Recent approaches have primarily designed sophisticated neural networks to capture highf-requency information for IMD tasks. However, these methods often overemphasize high-frequency information while overlooking the important role of low-frequency information in IMD tasks. To address this issue, we propose a Triple-Branch Frequency-Aware Network (TFNet), which includes an Information Separation Module (ISM), a Main Steam Branch (MSB), a Low-Frequency Learning Branch (LFL), a High-Frequency Learning Branch (HFL), and an Adaptive Aggregate Module (AAM) within a unified framework. Specifically, TFNet initially employs FSM to separate the manipulated image into RGB, high-frequency, and low-frequency components. Then, MSB, LFB, and HFB take the above components as input to learn the distinct features. Furthermore, the HFB is supervised with the boundary information to encourage the network to focus on the high-frequency information. Finally, the outputs of MSB, LFB, and HFB are sent to the ABM to adaptively aggregate features learned from the MSB, LFB, and HFB. Experiments on the CASIA, NIST, and Coverage datasets demonstrate the effectiveness of our TFNet.
作者机构：
[Sheng, Guo; She, Kang; Shan, Zhengping; Liu, Exian] College of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha 410004, China;[Peng, Yuchen] School of Engineering, Westlake University, Hangzhou, Zhejiang 310024, China;[Liu, Jianjun] The Key Laboratory for Micro/Nano-Optoelectronic Devices of Ministry of Education, School of Physics and Electronics, Hunan University, Changsha 410082, China
通讯机构：
[Exian Liu] C;College of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha 410004, China
摘要：
Topological photonic crystals (TPCs) are considered excellent optical sensor platform due to the robustness of light propagation, immunity to interference, and ease of integration. This paper theoretically proposes a topological valley photonic crystal (VPC) sensor based on second-order corner states and, to our knowledge, realizes for the first time topologically protected dual-parameter detection for refractive index (RI) and temperature. Two highly localized opological corner states (TCSs) are generated by splicing two types of VPCs with broken inversion symmetry. Owing to the very weak coupling between the two TCSs, they respond independently to RI and temperature variations, respectively. Results show that the RI sensitivity is around 653 nm/RIU over a wide range from 1.26 to 1.35, and the temperature sensitivity reaches 235 pm/°C over a broad range from 0 to 100 °C. Fascinatingly, the resonance peak's quality factor (Q) surpasses the order of 10(5) due to the high localization of TCS. Moreover, the TPC sensor maintains high-performance robustness by deliberately introducing 12 types of defects. This design has great potential in engineering compact, easy-integrated, and dual-parameter sensors in biomedical, chemical, and pharmaceutical fields.
摘要：
Multi-Unmanned Aerial Vehicle (UAV)-supported Mobile Edge Computing (MEC) can meet the computational requirements of tasks with high complexity and latency sensitivity to compensate for the lack of computational resources and coverage. In this paper, a multi-user and multi-UAV MEC networks is built as a two-tier UAV system in a task-intensive region where base stations are insufficient, with a centralized top-center UAV and a set of distributed bottom-UAVs providing computing services. The total energy consumption of the system is minimized by jointly optimizing the task offloading decision, 3D deployment of two-tier UAVs, the elevation angle of the bottom UAV, the number of UAVs, and computational resource allocation. To this end, an algorithm based on Differential Evolution and greedy algorithm with the objective of minimizing Energy Consumption (DEEC) is proposed in this paper. The algorithm uses a two-tier optimization framework where the upper tier uses a population optimization algorithm to solve for the location and elevation angle of the bottom UAV and the number of UAVs based on the actual ground equipment and the lower tier uses clustering and greedy algorithms to solve for the position of the top UAV, the offloading decision of the task, and the allocation of computational resources based on the results of the upper layer. The simulation results show that the algorithm effectively reduces the total energy consumption of the system while satisfying the task computation success rate and time delay.
Multi-Unmanned Aerial Vehicle (UAV)-supported Mobile Edge Computing (MEC) can meet the computational requirements of tasks with high complexity and latency sensitivity to compensate for the lack of computational resources and coverage. In this paper, a multi-user and multi-UAV MEC networks is built as a two-tier UAV system in a task-intensive region where base stations are insufficient, with a centralized top-center UAV and a set of distributed bottom-UAVs providing computing services. The total energy consumption of the system is minimized by jointly optimizing the task offloading decision, 3D deployment of two-tier UAVs, the elevation angle of the bottom UAV, the number of UAVs, and computational resource allocation. To this end, an algorithm based on Differential Evolution and greedy algorithm with the objective of minimizing Energy Consumption (DEEC) is proposed in this paper. The algorithm uses a two-tier optimization framework where the upper tier uses a population optimization algorithm to solve for the location and elevation angle of the bottom UAV and the number of UAVs based on the actual ground equipment and the lower tier uses clustering and greedy algorithms to solve for the position of the top UAV, the offloading decision of the task, and the allocation of computational resources based on the results of the upper layer. The simulation results show that the algorithm effectively reduces the total energy consumption of the system while satisfying the task computation success rate and time delay.
期刊：
JOURNAL OF SUPERCOMPUTING,2025年81(1):1-21 ISSN：0920-8542
通讯作者：
Xuyu Xiang
作者机构：
[Peng, Jianting; Xiang, Xuyu; Qin, Jiaohua; Tan, Yun] College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, China
通讯机构：
[Xuyu Xiang] C;College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, China
摘要：
Cross-modal retrieval can break through the limitations of modalities and carry out information retrieval across data of different modalities to meet the needs of users in obtaining multi-modal correlation retrieval. Cloud computing has the advantages of high efficiency and low cost, but data security hinders its development. While cloud computing offers high efficiency and cost-effectiveness, concerns surrounding data security impede its full potential. Privacy-preserving cross-modal retrieval emerges as a viable solution, catering to users’ demands for efficient retrieval while safeguarding data confidentiality. However, a major challenge still exists in this field: how to bridge the inherent semantic gap within heterogeneous and chaotic information. To address this challenge, this paper proposes dual-branch networks for privacy-preserving cross-modal retrieval in cloud computing. Firstly, a dual-branch feature extraction network of encrypted image-text is constructed, enhancing the extraction of meaningful features from encrypted data. Secondly, a cross-modal alignment method is designed to eliminate the heterogeneous gap between different modalities through the alignment within and between modalities. Finally, to fully exploit the storage and computing advantages of cloud computing, both encrypted data and the cross-modal feature extractor are deployed to the cloud. Leveraging the dynamic update capabilities of cloud-stored encrypted data enables continuous model refinement, enhancing retrieval accuracy while reducing the storage and computational burdens on data owners. Extensive experiments conducted on the publicly available benchmark image-text dataset Wikipedia indicate that, compared to existing methods, our approach achieves improvements of 5.4%, 1%, 1.6%, and 20.1% in the four metrics of image-to-text (i2t), text-to-image (t2i), image-to-all (i2all), and text-to-all (t2all), respectively.
摘要：
With the continuous development of deep learning (DL), the task of multimodal dialog emotion recognition (MDER) has recently received extensive research attention, which is also an essential branch of DL. The MDER aims to identify the emotional information contained in different modalities, e.g., text, video, and audio, and in different dialog scenes. However, the existing research has focused on modeling contextual semantic information and dialog relations between speakers while ignoring the impact of event relations on emotion. To tackle the above issues, we propose a novel dialog and event relation-aware graph convolutional neural network (DER-GCN) for multimodal emotion recognition method. It models dialog relations between speakers and captures latent event relations information. Specifically, we construct a weighted multirelationship graph to simultaneously capture the dependencies between speakers and event relations in a dialog. Moreover, we also introduce a self-supervised masked graph autoencoder (SMGAE) to improve the fusion representation ability of features and structures. Next, we design a new multiple information Transformer (MIT) to capture the correlation between different relations, which can provide a better fuse of the multivariate information between relations. Finally, we propose a loss optimization strategy based on contrastive learning to enhance the representation learning ability of minority class features. We conduct extensive experiments on the benchmark datasets, Interactive Emotional Dyadic Motion Capture (IEMOCAP) and Multimodal EmotionLines Dataset (MELD), which verify the effectiveness of the DER-GCN model. The results demonstrate that our model significantly improves both the average accuracy and the $F1$ value of emotion recognition. Our code is publicly available at https://github.com/yuntaoshou/DER-GCN.
通讯机构：
[Liu, EX ] C;Cent South Univ Forestry & Technol, Coll Comp & Informat Engn, Changsha 410000, Peoples R China.
关键词：
Sensors;Photonic band gap;Couplings;Lattices;Refractive index;Biosensors;Band structures;Refractive index (RI) sensor;topological corner state (TCS);topological photonic crystals (TPC);ultrahigh quality factor
摘要：
Owing to the flexible light control and the immunity to defects and impurities, topological photonic crystals (TPC) are considered an excellent platform in the engineering of optoelectronic devices. Herein, four TPC refractive index (RI) sensors are theoretically proposed based on the su-Schrieffer-Heeger (SSH) model. Zero-dimensional topological corner states (TCS) and 1-D topological edge states (TES) are demonstrated by the band structure and the Zak phases. Four TPC sensors are based on single topological microcavity (TCS formed by two TES and four TES), trivial corner states (TRCS)-TES coupling, and TCS-TES coupling. The TCS-based TPC sensors exhibit an ultrahigh quality factor (Q) of $\gt 10^{{6}}$ . Moreover, TPC sensor based on TCS-TES coupling remain a topological robustness with >99.4% sensitivity quality of reproduction by artificially introducing eight-type defects. Notably, the design based on TCS-TES coupling presents a sensitivity of 413.76 nm/RIU and a Q value of $2.8\times 10^{{6}}$ in a broad RI range of 1.00–1.50. This work not only provides a new way of 2-D photonic crystals (PC) sensor applications but also offers a guidance for the design of TPC sensors.
摘要：
The popularity of smart devices has made it more convenient for users to take screen shots, but it has also made it easier to take clandestine shots, resulting in compromised and untraceable information. Therefore, this paper introduces a screen-shooting robust watermarking algorithm based on the pattern complexity just noticeable difference (PC-JND) model. This approach involves the utilisation of local binary patterns (LBP) for block filtering based on texture complexity in the original image. Stable feature blocks are selected and processed using the speeded-up robust features (SURF) algorithm to extract key feature points, defining them as the watermark embedding regions. Finally, the watermark is embedded in the integer wavelet domain's HH sub-band, guided by the PC-JND model. Experimental results demonstrate that this algorithm not only significantly improves the visual quality of images in a shorter embedding time but also exhibits enhanced robustness against screen captures from various angles.
期刊：
BRIEFINGS IN BIOINFORMATICS,2025年26(5) ISSN：1467-5463
通讯作者：
Chen, YF;Lu, AP
作者机构：
[Chen, Yifan; Fu, Xiangzheng] Cent South Univ Forestry & Technol, Coll Comp & Informat Engn, Inst Artificial Intelligence Applicat, 498 Shaoshan South Rd,Tianxin Dist, Changsha 410004, Hunan, Peoples R China.;[Lu, Aiping; Fu, Xiangzheng] Hong Kong Baptist Univ, Sch Chinese Med, Kowloon Tong, Kowloon, 15 Baptist Univ Rd, Hong Kong 999077, Peoples R China.;[Chen, Haowen] Hunan Univ, Coll Sci & Elect Engn, 2 Lushan South Rd, Changsha 410082, Hunan, Peoples R China.;[Peng, Li] Hunan Univ Sci & Technol, Coll Comp Sci & Engn, 1 Taoyuan Rd, Xiangtan 411201, Hunan, Peoples R China.;[Rong, Mingqiang] Hunan Normal Univ, Coll Life Sci, Natl & Local Joint Engn Lab Anim Peptide Drug Dev, 36 Lushan Rd, Changsha 410081, Hunan, Peoples R China.
通讯机构：
[Chen, YF ] C;[Lu, AP ] H;Cent South Univ Forestry & Technol, Coll Comp & Informat Engn, Inst Artificial Intelligence Applicat, 498 Shaoshan South Rd,Tianxin Dist, Changsha 410004, Hunan, Peoples R China.;Hong Kong Baptist Univ, Sch Chinese Med, Kowloon Tong, Kowloon, 15 Baptist Univ Rd, Hong Kong 999077, Peoples R China.;Changsha Med Univ, Sch Informat Engn, 1501 Leifeng Rd, Changsha 410219, Peoples R China.
关键词：
AUC-maximization;TCR-epitope binding;graph regularization;protein language models
摘要：
T-cell receptor (TCR)-epitope binding prediction is critical for immunotherapies but remains challenged by sparse interaction networks and severe class imbalance in training data. Current graph neural network (GNN) approaches for predicting TCR-epitope binding (TEB) fail to address two key limitations: over-smoothing during message propagation in sparse TCR-epitope graphs and biased predictions toward dominant epitope-TCR pairs. Here, we present GRAPE (Graph-Regularized Attentive Protein Embeddings), a framework unifying spectral graph regularization and imbalance-aware learning. GRAPE first leverages protein language models (ESM-2) to generate evolutionary-informed TCR/epitope embeddings, constructing a topology-aware interaction graph. To mitigate over-smoothing, we introduce spectral graph regularization, explicitly constraining node feature smoothness to preserve discriminative patterns in sparse neighborhoods. Simultaneously, a dynamic edge reweighting module prioritizes unobserved TCR-epitope edges during graph propagation, coupled with a differentiable area under the ROC curve-maximization objective that directly optimizes for imbalance resilience. Extensive benchmarking on public datasets demonstrates that GRAPE significantly outperforms state-of-the-art methods in TEB prediction. This work establishes GRAPE as a robust framework for elucidating TCR-epitope interactions, with broad applications in immunology research and therapeutic design.
期刊：
International Journal of Autonomous and Adaptive Communications Systems,2025年18(4):341-356 ISSN：1754-8632
通讯作者：
Qin, JH
作者机构：
[Qianyi Liu; Jiaohua Qin; Xuyu Xiang; Yun Tan] College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, 410004, China
摘要：
The surge in online shopping has heightened the demand for interactive fashion design retrieval. Existing methods, however, exhibit imperfections in attribute segmentation, attributed to the specificity of clothing attributes. The attention region often encounters multiple attributes overlapping, causing changes in one attribute to affect irrelevant ones, resulting in poor retrieval accuracy. This paper addresses this challenge by proposing a deep neural network for fashion retrieval based on multi-attention attribute manipulation. In this approach, the feature extraction module sifts the extracted features to obtain an overall description of the clothing image by adding ESE-NAM combined attention modules to the VoVNet network block. The attribute decoding module utilises one-hot coding and feature mapping to subdivide the attribute features, obtaining more independent local detail features for refined attribute image retrieval with a focus on details. Experimental results show that the proposed network surpasses existing networks with an overall accuracy increase of more than 4% points, particularly with the feature extraction module demonstrating an accuracy boost of over 6% points.
作者机构：
[Sun, Haiwen; Tang, Chunling] Cent South Univ Forestry & Technol, Coll Econ & Management, Changsha 410004, Hunan, Peoples R China.;[Gao, Ping] Univ Manchester, Global Dev Inst, Oxford Rd, Manchester M13 9PL, England.;[Zhou, Guoxiong] Cent South Univ Forestry & Technol, Coll Comp & Informat Engn, Changsha 410004, Hunan, Peoples R China.
通讯机构：
[Tang, CL ] C;Cent South Univ Forestry & Technol, Coll Econ & Management, Changsha 410004, Hunan, Peoples R China.
关键词：
Digital economy;Energy consumption;Carbon emissions;Non-linear relationships;Spatial spillover effect
摘要：
The intensifying challenge of global climate change has made accelerating energy conservation and emission reduction an urgent global imperative. As one of the world’s largest carbon emitters, China plays a pivotal role in global decarbonization efforts. The digital economy, emerging as a key driver of China’s economic transformation, provides novel pathways for advancing carbon reduction. This paper employs kernel density estimation and ArcGIS 10.8 to analyze the spatiotemporal dynamics of the digital economy and urban carbon emissions in China. Using panel data from 278 prefecture-level cities, the study applies fixed-effects models, mediation effect models, and spatial Durbin models to explore the mechanisms and spatial impacts of the digital economy on carbon reduction. The findings reveal that: (1) the development of the digital economy exerts a significant “inverted U-shaped” influence on urban carbon emission reduction; (2) energy consumption intensity is the critical mechanism underlying the nonlinear relationship between the digital economy and carbon emissions; (3) the digital economyʼs impact on carbon emissions exhibits spatial spillover effects, following a similar “inverted U-shaped” trajectory. These results contribute valuable empirical insights into the dual objectives of digital economic growth and carbon emission reduction, offering policymakers guidance on leveraging digitalization to achieve sustainable and coordinated regional development.
The intensifying challenge of global climate change has made accelerating energy conservation and emission reduction an urgent global imperative. As one of the world’s largest carbon emitters, China plays a pivotal role in global decarbonization efforts. The digital economy, emerging as a key driver of China’s economic transformation, provides novel pathways for advancing carbon reduction. This paper employs kernel density estimation and ArcGIS 10.8 to analyze the spatiotemporal dynamics of the digital economy and urban carbon emissions in China. Using panel data from 278 prefecture-level cities, the study applies fixed-effects models, mediation effect models, and spatial Durbin models to explore the mechanisms and spatial impacts of the digital economy on carbon reduction. The findings reveal that: (1) the development of the digital economy exerts a significant “inverted U-shaped” influence on urban carbon emission reduction; (2) energy consumption intensity is the critical mechanism underlying the nonlinear relationship between the digital economy and carbon emissions; (3) the digital economyʼs impact on carbon emissions exhibits spatial spillover effects, following a similar “inverted U-shaped” trajectory. These results contribute valuable empirical insights into the dual objectives of digital economic growth and carbon emission reduction, offering policymakers guidance on leveraging digitalization to achieve sustainable and coordinated regional development.
关键词：
RDH;reversible data hiding;prediction error;Huffman coding;encrypted images;extended run-length coding
摘要：
To reduce prediction errors and create more room for embedding data, the paper proposes a reversible data hiding (RDH) in encrypted images scheme based on histogram shifting and prediction error block coding. Firstly, the histogram of the prediction error image is shifted according to the signs of prediction errors. Next, the prediction error plane is partitioned into uniformly sized blocks, and these blocks are labelled as three types: an all-zero block, a block containing only one 1, and a block containing more than one 1. These three types of blocks are compressed using labelling, binary tree coding, and Huffman coding, respectively. To better compress the label map, an improved extended run-length coding is proposed. Finally, the image is secured by encryption and the secret data is hidden within it. The experimental results indicate a significant improvement in the embedding rate of the scheme compared to other schemes.
期刊：
Biomedical Signal Processing and Control,2025年107:107815 ISSN：1746-8094
通讯作者：
Genhua Liu<&wdkj&>Guoxiong Zhou
作者机构：
["Yang, Yixin; Sun, Lixiang; Tang, Zhiwen; Liu, Genhua; Zhou, Guoxiong; Li, Lin] School of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha 410004, China;[Cai, Weiwei] School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China;[Li, Liujun] Department of Civil, Architectural and Environmental Engineering, Missouri University of Science & Technology, Rolla 65401, USA;[Chen, Lin] Zhuzhou Sany Aier Eye Hospital, Zhuzhou 412002, China;[Hu, Linan"] Department of Radiology, Zhuzhou Central Hospital, Zhuzhou 412001, China
通讯机构：
[Genhua Liu; Guoxiong Zhou] S;School of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha 410004, China
摘要：
In 2013, an estimated 64 million people between the ages of 40 and 80 were suffering from eye disease. By 2020, that number had climbed to 76 million. It is predicted that by 2040, there will be a staggering 111.8 million glaucoma patients worldwide. Segmentation of blood vessels in retinal images can be used to investigate many diseases, and the complexity of the blood vessels and the variable conditions inside the retina pose a high challenge for accurate segmentation. Therefore, a competing fusion segmentation network (TAOD −CFNet) with a trumpet-like attention mechanism and optic disc gradient adjustment algorithm for retinal blood vessel segmentation. First, an optic disc gradient adjustment algorithm (ODGA) is proposed, which designs dual threshold weights for accurate localization and optimization of optic disc areas. Then, a competing fusion block (CFB) is proposed to improve the feature dissimilarity between the arteriovenous vascular sensitive area and the interference area. Finally, a Trumpet Attention Mechanism (TAM) is proposed to enhance the edge features of fine and peripheral blood vessels. TAOD-CFNet outperforms ten SOTA methods in ten-fold cross-validation, with IOU, F1-Score, Dice, Jaccard, ACC and MCC metrics reaching 83.28%, 89.41%, 84.28%, 80.35%, 96.94% and 88.81%. To demonstrate the generalization performance of the model, TAOD-CFNet outperforms ten SOTA image segmentation methods on six retinal image datasets (DRIVE, CHASEDB1, STARE, HRF, IOSTAR, and LES). The experimental results proved that TAOD-CFNet possesses better segmentation performance and can assist clinicians in determining the condition of retinopathy patients.
In 2013, an estimated 64 million people between the ages of 40 and 80 were suffering from eye disease. By 2020, that number had climbed to 76 million. It is predicted that by 2040, there will be a staggering 111.8 million glaucoma patients worldwide. Segmentation of blood vessels in retinal images can be used to investigate many diseases, and the complexity of the blood vessels and the variable conditions inside the retina pose a high challenge for accurate segmentation. Therefore, a competing fusion segmentation network (TAOD −CFNet) with a trumpet-like attention mechanism and optic disc gradient adjustment algorithm for retinal blood vessel segmentation. First, an optic disc gradient adjustment algorithm (ODGA) is proposed, which designs dual threshold weights for accurate localization and optimization of optic disc areas. Then, a competing fusion block (CFB) is proposed to improve the feature dissimilarity between the arteriovenous vascular sensitive area and the interference area. Finally, a Trumpet Attention Mechanism (TAM) is proposed to enhance the edge features of fine and peripheral blood vessels. TAOD-CFNet outperforms ten SOTA methods in ten-fold cross-validation, with IOU, F1-Score, Dice, Jaccard, ACC and MCC metrics reaching 83.28%, 89.41%, 84.28%, 80.35%, 96.94% and 88.81%. To demonstrate the generalization performance of the model, TAOD-CFNet outperforms ten SOTA image segmentation methods on six retinal image datasets (DRIVE, CHASEDB1, STARE, HRF, IOSTAR, and LES). The experimental results proved that TAOD-CFNet possesses better segmentation performance and can assist clinicians in determining the condition of retinopathy patients.
作者：
Feng Shi;Yicong Zhu;Guangwei Wu;Jingyi Liu;Jianxin Wang
作者机构：
[Feng Shi; Yicong Zhu; Jingyi Liu; Jianxin Wang] School of Computer Science and Engineering, Central South University, 410083, Changsha, People’s Republic of China;[Guangwei Wu] College of Computer and Information Engineering, Central South University of Forestry and Technology, 410004, Changsha, People’s Republic of China
会议名称：
Computing and Combinatorics: 31st International Computing and Combinatorics Conference, COCOON 2025, Chengdu, China, August 15–17, 2025, Proceedings, Part II
摘要：
Within the paper, we study several variants of the decision problem, Scheduling with precedence constraints and time windows, denoted by P ∣ p r e c , r i , d i ∣ ⋆ , and present improved fixed-parameter algorithms parameterized by the maximum processing time p max and the maximum number μ of overlapping time windows, defined as μ = max t ∈ N | { i ∈ S ∣ r i ≤ t < d i } | . Firstly, we propose an algorithm for P ∣ p r e c , r i , d i ∣ ⋆ with time complexity O ( ( p max + 2 ) μ p max n 3 ) , where n is the number of tasks. This significantly improves the previously best-known algorithm with time complexity O ( p max 2 μ · 16 μ μ · n 3 ) . Then for the unit processing time case P ∣ p r e c , p i = 1 , r i , d i ∣ ⋆ , we further develop an algorithm with time complexity O ( 2 μ μ m n 3 ) , where m is the number of machines, improving the previously best-known algorithm with time complexity O ( 16 μ n 4 ) . Finally, we extend the two algorithms to the typed machine setting, solving P ∣ M j ( t y p e ) , p r e c , r i , d i ∣ ⋆ and P ∣ M j ( t y p e ) , p r e c , p i = 1 , r i , d i ∣ ⋆ , with time complexities O ( ( p max + 2 ) μ p max n 3 ) and O ( 2 μ μ m k n 3 ) , respectively, where k is the number of machine types.
摘要：
Tomatoes are an important global crop, and automating the segmentation of leaf diseases is essential for agricultural security. Effective segmentation of these diseases is vital for timely intervention, which can significantly enhance crop yield and reduce pesticide usage. However, challenges such as background interference, tiny diseases, and blurred disease edges pose immense obstacles to the segmentation of tomato leaf diseases. To address these issues effectively, we propose a fusion adversarial segmentation network for tomato disease segmentation named FATDNet. Firstly, to eliminate background interference effectively, we introduce a dual-path fusion adversarial algorithm (DFAA). This algorithm employs parallel dual-path convolution to extract features of leaf disease regions, merging and adversarially processing complex background noise and disease features. Secondly, to enhance the feature extraction of small lesion regions, we employ a multi-dimensional attention mechanism (MDAM). This mechanism allocates weights in both horizontal and vertical directions, subsequently calculating weights in different channels of the feature map. This enhances the dispersion of semantic information through the adoption of diverse weight calculation strategies. Furthermore, to improve the model’s ability to extract features at the edges of leaf diseases, we introduce a Gaussian weighted edge segmentation module (GWESM). This module calculates weight distribution through a Gaussian-weighted function, guiding the model to highlight features at different scales and reduce information loss caused by pooling operations. To demonstrate the superiority of FATDNet, we conduct comparative experiments using a self-built dataset and a public dataset. Experimental results show that FATDNet outperforms nine state-of-the-art segmentation networks. This validates that FATDNet provides a reliable solution for the automated segmentation of tomato leaf diseases.
Tomatoes are an important global crop, and automating the segmentation of leaf diseases is essential for agricultural security. Effective segmentation of these diseases is vital for timely intervention, which can significantly enhance crop yield and reduce pesticide usage. However, challenges such as background interference, tiny diseases, and blurred disease edges pose immense obstacles to the segmentation of tomato leaf diseases. To address these issues effectively, we propose a fusion adversarial segmentation network for tomato disease segmentation named FATDNet. Firstly, to eliminate background interference effectively, we introduce a dual-path fusion adversarial algorithm (DFAA). This algorithm employs parallel dual-path convolution to extract features of leaf disease regions, merging and adversarially processing complex background noise and disease features. Secondly, to enhance the feature extraction of small lesion regions, we employ a multi-dimensional attention mechanism (MDAM). This mechanism allocates weights in both horizontal and vertical directions, subsequently calculating weights in different channels of the feature map. This enhances the dispersion of semantic information through the adoption of diverse weight calculation strategies. Furthermore, to improve the model’s ability to extract features at the edges of leaf diseases, we introduce a Gaussian weighted edge segmentation module (GWESM). This module calculates weight distribution through a Gaussian-weighted function, guiding the model to highlight features at different scales and reduce information loss caused by pooling operations. To demonstrate the superiority of FATDNet, we conduct comparative experiments using a self-built dataset and a public dataset. Experimental results show that FATDNet outperforms nine state-of-the-art segmentation networks. This validates that FATDNet provides a reliable solution for the automated segmentation of tomato leaf diseases.
作者：
Lin Tan;Songtao Guo;Zhufang Kuang;Pengzhan Zhou;Mingyan Li
期刊：
IEEE Transactions on Wireless Communications,2025年:1-1 ISSN：1536-1276
作者机构：
[Lin Tan; Songtao Guo; Pengzhan Zhou; Mingyan Li] Key Laboratory of Dependable Service Computing in Cyber-Physical-Society (Ministry of Education), and College of Computer Science, Chongqing University, Chongqing, China;[Zhufang Kuang] School of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha, China
摘要：
Low Earth Orbit (LEO) satellite networks hold great promise in the field of wireless communication due to their global coverage. However, the long communication distances and massive data computations present significant challenges for current satellite networks. To overcome these barriers, we propose SkyLink, a universal Integrated Ground-Air-Space Collaborative Edge Computing system that leverages horizontal collaboration among aerial platforms (AirXs) as well as vertical collaboration among Ground-Air-Space. We propose a bi-level optimization framework based on a Multi-Agent Twin Delayed Deep Deterministic policy gradient (MATD3) with Hybrid Action Space and constructe a latent representation space for each agent to allow the agent to learn the latent policy. This enabling each AirX to act as an agent and autonomously optimize its hybrid action decisions to improve system efficiency in real-time based on the dynamic network environment, a capability not achievable by conventional DRL methods. This includes continuous optimization variables such as AirX deployment (location changes) and resource allocation, as well as discrete optimization variables for collaborative task offloading decisions. Extensive experiments against state-of-the-art algorithms (e.g., MADDPG, QMIX) demonstrate that the proposed system improves energy efficiency by 27.2% and task completion rate by 6.8% compared to traditional Integrated Ground-Air-Space (IG) Network.
