Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2024 Vol.37 Issue.7, Published 2024-07-25 Papers and Reports    Researches and Applications        Papers and Reports 571 Event-Driven Story Writing Based on Three-Act Structural Chain-of-Thought and Semantic Self-Consistency HUANG Yuxin, ZHAO Yuan, YU Zhengtao, WU Lei, MA Jiushun Event-driven story writing aims to create coherent stories that conform to event content based on limited background and event information. However, existing methods often suffer from semantic incoherence and plot conflicts due to insufficient reasoning about complex event relationships. To address these problems, a method for event-driven story writing based on three-act structural chain-of-thought and semantic self-consistency is proposed in this paper. Before generating the story, diverse story examples are selected to enable the model to learn different storytelling styles. During the story generation, a chain-of-thought is designed based on three-act structure of setup, confrontation and resolution, guiding the model to reasonably plan the story content and avoid plot inconsistencies. After the story is generated, semantic self-consistency is introduced to simulate the writer's deliberation process, selecting the most semantically consistent, coherent and relevant story from multiple generated versions. Experiments show that the proposed method improves BLEU-4 and BERTScore metrics and demonstrates certain advantages in human evaluations as well. 2024 Vol. 37 (7): 571-583 [Abstract] ( 386 ) [HTML 1KB] [ PDF 967KB] ( 510 ) 584 Triadic Concept Construction Method Based on Candidate Set WANG Xiao, WEI Ling, ZHANG Qin, QI Bin As an extension of formal concept analysis, triadic concept analysis is a theory for analyzing three-dimensional data. The acquisition of triadic concepts is one of the key issues in triadic concept analysis. A triadic concept construction method based on candidate set is proposed. Firstly, the regular triadic context and the purified triadic context are defined, and properties of these two triadic contexts are studied. Secondly, it is proven that the extent set of all formal concepts of the formal context induced by the triadic context contains the extent set of all triadic concepts of triadic context. Then, by defining an extent candidate set, a method for constructing triadic concepts using the extent candidate set is presented to speed up the acquisition of triadic concepts. Moreover, the feasibility and completeness of obtaining triadic concepts based on this construction method are proven, and this method is extended to two other types of formal contexts induced by the triadic context. Finally, an algorithm for constructing triadic concepts based on the candidate set is presented, and experimental results demonstrate superior performance of the proposed algorithm. 2024 Vol. 37 (7): 584-596 [Abstract] ( 191 ) [HTML 1KB] [ PDF 827KB] ( 345 ) 597 Multi-evolutionary Features Based Link Prediction Algorithm for Social Network HE Yulin, LAI Junlong, CUI Laizhong, HUANG Zhexue, YIN Jianfei Social network link prediction aims to predict future link relationships based on known network information, in which there are important applications for recommender systems and co-authorship networks. However, existing link prediction algorithms often ignore multi-evolutionary features of social networks and have high training time complexity, limiting their execution efficiency and application performance. To address these problems, a multi-evolutionary features based link prediction algorithm for social network(MEF-LP) is proposed. Firstly, a simple and efficient time extreme learning machine model is designed to transfer and aggregate the temporal information of social network snapshot sequences, using gated networks and extreme learning machine self-encoders. Secondly, multiple multilayer extreme learning machines are constructed to map temporal features from multiple perspectives, mining different evolutionary features of social networks and ultimately integrating them into comprehensive evolutionary features. Finally, the extreme learning machine-based classifiers are utilized to complete the link prediction. Experiments on six real social networks show that MEF-LP can reasonably learn the multi-evolution features of social networks and achieve better prediction performance. 2024 Vol. 37 (7): 597-612 [Abstract] ( 231 ) [HTML 1KB] [ PDF 3257KB] ( 414 ) 613 Collaborative Brain Age Prediction Algorithm Based on Multimodal Fuzzy Feature Fusion WANG Jing, DING Weiping, YIN Tao, JU Hengrong, HUANG Jiashuang Deep neural networks can be trained to predict age from brain image and the predicted brain age serves as a biomarker for identifying diseases associated with aging. Traditional brain age prediction methods tend to rely on unimodal image data, whereas multimodal data can provide more comprehensive information and improve prediction accuracy. However, existing multimodal fusion methods often fail to fully exploit the correlations and complementarities between different modalities. To overcome these challenges, a collaborative brain age prediction algorithm based on multimodal fuzzy feature fusion(CMFF) is proposed. Fuzzy fusion module and multimodal collaborative convolution module are designed to effectively utilize the correlation and complementary information between the multimodal information. Firstly, feature tensors are extracted from multimodal brain images by a convolutional neural network, and are integrated into a global feature tensor via radial joins. Then, the fuzzy fusion module is employed to learn the fuzzified features, and these features are applied to the multimodal collaborative convolutional module to enhance the complementary information of these features through modality-specific convolutional layers. Finally, the age prediction regression task is performed based on the gender information and the fuzzy collaborative processed features to obtain an accurate predicted age. Experimental results on SRPBS multi-disorder MRI dataset demonstrate the superior performance of CMFF. 2024 Vol. 37 (7): 613-625 [Abstract] ( 224 ) [HTML 1KB] [ PDF 1074KB] ( 375 ) Researches and Applications 626 Image Super-Resolution Reconstruction Method Based on Lightweight Symmetric CNN-Transformer WANG Tingwei, ZHAO Jianwei, ZHOU Zhenghua To address the issues of large parameter sizes and high training cost in existing image super-resolution reconstruction methods based on Transformer, an image super-resolution reconstruction method based on lightweight symmetric CNN-Transformer is proposed. Firstly, a symmetric CNN-Transformer block is designed using weight sharing, and the information extracted from the upper and lower branches is fully integrated through channel attention block to improve the ability of the network to capture and utilize both local and global features. Meanwhile, based on the depthwise separable convolution and the calculation of the self-attention cross-channel covariance matrix, the number of parameters in Transformer is effectively decreased, as well as calculation cost and memory consumption. Secondly, a high-frequency enhancement residual block is introduced into the network to further focus on the texture and detail information in the high-frequency area. Finally, the selection of the best activation function for generating the self-attention in Transformer is explored. Experimental analysis demonstrates that GELU function can better promote feature aggregation and improve network performance. Experimental results show that the proposed method effectively reconstructs richer textures and details of the image while maintaining the lightweight of the network. 2024 Vol. 37 (7): 626-637 [Abstract] ( 413 ) [HTML 1KB] [ PDF 1926KB] ( 538 ) 638 Image Generation Method for Cognizing Image Attribute Features from the Perspective of Disentangled Representation Learning CAI Jianghai, HUANG Chengquan, WANG Shunxia, LUO Senyan, YANG Guiyan, ZHOU Lihua In the field of generative artificial intelligence, the research of disentangled representation learning further promotes the development of image generation methods. However, existing disentanglement methods pay more attention to low-dimensional representation of image generation, ignoring inherent interpretable factors of the target variation image. This oversight results in generated image being susceptible to the influence of other irrelevant attribute features. To address this issue, an image generation method for cognizing image attribute features from the perspective of disentangled representation learning is proposed. Firstly, candidate traversal directions for the target variation image are obtained by training, starting from the latent space of the generative model. Secondly, an unsupervised semantic decomposition strategy is constructed, and the interpretable directions embedded in the latent space are jointly discovered based on the direction of candidate traversals. Finally, a contrast simulator and a variation space are constructed using disentangled encoders and contrastive learning. Consequently, the disentangled representations of the target variation image are extracted from the interpretable directions and the image is generated. Extensive experiments on five popular disentanglement datasets demonstrate the superior performance of the proposed method. 2024 Vol. 37 (7): 638-651 [Abstract] ( 210 ) [HTML 1KB] [ PDF 4136KB] ( 405 ) 652 Depth-Reshaping Based Aerial Object Detection Enhanced Network FU Tianyi, YANG Benyi, DONG Hongbin, DENG Baosong To address the issues of complex background interference, loss of fine details in small objects and the high demand for detection efficiency in aerial image object detection, a depth-reshaping enhanced network(DR-ENet) is proposed. Firstly, the traditional downsampling methods are replaced by spatial depth-reshaping techniques to reduce information loss during feature extraction and enhance the ability of the network to capture details. Then, a deformable spatial pyramid pooling method is designed to enhance the adaptability of network to object shape variations and its ability to recognize in complex backgrounds. Simultaneously, an attention decoupling detection head is proposed to enhance the learning effectiveness for different detection tasks. Finally, a small-scale aerial dataset , PORT, is constructed to simultaneously consider the characteristics of dense small objects and complex backgrounds. Experiments on three public aerial datasets and PORT dataset demonstrate that DR-ENet achieves performance improvement, proving its effectiveness and high efficiency in aerial image object detection. 2024 Vol. 37 (7): 652-662 [Abstract] ( 246 ) [HTML 1KB] [ PDF 2486KB] ( 365 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn