Article
Computer Science, Artificial Intelligence
Andres Romero, Juan Leon, Pablo Arbelaez
Summary: We propose a novel convolutional neural network approach for the fine-grained recognition problem of multi-view dynamic facial action unit detection. Our approach leverages recent advances in large-scale object recognition and considers both shared and independent representations as well as different CNN architectures. Extensive experiments on the FERA 2017 Challenge demonstrate the effectiveness and superiority of our approach.
IMAGE AND VISION COMPUTING
(2022)
Article
Agriculture, Multidisciplinary
Ruotong Yang, Zikang Chen, Huanliang Xu, Mingxia Shen, Pinghua Li, Tomas Norton, Mingzhou Lu
Summary: This study developed an automatic method for recognizing the rooting actions of prepartum sows using computer vision. The method includes object detection, optical flow estimation, and support vector machine classification. The test results showed that the method can accurately identify whether a video contains rooting actions and can be used to recognize the nest-building behavior and predict the delivery time of sows.
COMPUTERS AND ELECTRONICS IN AGRICULTURE
(2023)
Article
Engineering, Electrical & Electronic
Jun Wang
Summary: We propose an action recognition method based on Riemannian manifold and adaptive weighted feature fusion. The method eliminates background disturbance and focuses on the moving region of the target, and introduces the Riemannian manifold distance feature to capture the change rate of appearance and motion. Experimental results show that the method has significant advantages in action recognition.
JOURNAL OF ELECTRONIC IMAGING
(2022)
Article
Computer Science, Software Engineering
Manisha Mudgal, Deepika Punj, Anuradha Pillai
Summary: Research in the field of image processing and computer vision for recognition of suspicious activity is actively growing, with a focus on surveillance systems and violence activities such as hitting, slapping, punching. This research requires large human action datasets and proposes a method using Gaussian Mixture Model to model violence actions.
JOURNAL OF WEB ENGINEERING
(2021)
Article
Mathematics
Dong Chen, Tao Zhang, Peng Zhou, Chenyang Yan, Chuanqi Li
Summary: This paper proposes a novel pose-based action representation method called Optical Flow Pose Image (OFPI) to fully leverage the spatial and temporal information of skeletal data. The OFPI representation achieved 98.3% and 94.2% accuracy on the KTH and JHMDB datasets, respectively, surpassing other methods and demonstrating the utility and potential of this algorithm for skeleton-based action recognition research.
Article
Computer Science, Hardware & Architecture
Limin Xia, Wentao Ma
Summary: This study proposes an optical flow-based physical feature-driven action recognition framework, which calculates the original optical flow field, develops joint action relevance, and processes local spatial-temporal thermal diffusion to obtain a more stable flow field. A feature descriptor is designed to consider divergence, curl, and gradient features of the flow field, and Fisher vectors are adopted for encoding descriptors for classification. Experimental results on HMDB51, KTH, and UCF101 datasets demonstrate that the proposed framework can accurately recognize actions in complex backgrounds, outperforming existing methods.
JOURNAL OF SUPERCOMPUTING
(2021)
Article
Engineering, Civil
Wenjie Song, Shixian Liu, Ting Zhang, Yi Yang, Mengyin Fu
Summary: This paper proposes a vehicle taillight recognition method based on action-state joint learning, which achieves practical results in complex actual scenes by considering taillight state features and time series features. The method uses vehicle tracking sequences as input, identifies the action features of brake lights and turn signals using a CNN-LSTM model, and infers the continuous taillight state by analyzing higher-level features extracted through semantic segmentation.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2022)
Article
Computer Science, Information Systems
Andre de Souza Brito, Marcelo Bernardes Vieira, Saulo Moraes Villela, Hemerson Tacon, Hugo de Lima Chaves, Helena de Almeida Maia, Darwin Ttito Concha, Helio Pedrini
Summary: The study proposes a method based on a multi-stream architecture and weighted voting for recognizing human actions in videos. By introducing a new Optical Flow Rhythm stream and using a new weighted average fusion method to combine different feature streams, optimizing classifier weights, experiments show that the method performs comparably to current state-of-the-art approaches.
JOURNAL OF VISUAL COMMUNICATION AND IMAGE REPRESENTATION
(2021)
Article
Computer Science, Information Systems
Soolmaz Abbasi, Mehdi Rezaeian
Summary: This paper proposes a correlation filter-based tracker that combines multiple features to extract texture and motion model information, capable of handling occlusion and out-of-view in tracking. Additionally, the proposed method utilizes adaptive thresholding and similarity transformation for robust estimation and prediction of target scale, rotation, and translation changes.
MULTIMEDIA TOOLS AND APPLICATIONS
(2021)
Article
Engineering, Civil
Xudong Fan, Wei Zhao
Summary: This paper proposes a robust license plate detection network and a segmentation-free network for accurate license plate recognition and rectification under complex capture scenarios. Experimental results demonstrate the good performance of the system.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2022)
Article
Computer Science, Information Systems
Qili Zeng, M. Ozan Tezcan, Janusz Konrad
Summary: In this paper, a simple and flexible Dynamic Equilibrium Module (DEM) for video modeling through adaptive Eulerian motion manipulation is proposed. The module can be inserted into backbone networks to reduce the impact of temporal variations on video modeling and learn spatio-temporal representations with higher robustness. Performance gains are demonstrated in R3D and R(2+1)D models on Kinetics-400, UCF-101, and HMDB-51 datasets.
Article
Computer Science, Artificial Intelligence
Wei Lin, Xinghao Ding, Yue Huang, Huanqiang Zeng
Summary: This paper proposes a self-supervised video-based action recognition method (VARD), which extracts the principal information of actions by combining visual and semantic information for more accurate action representation. By disturbing positive samples in both visual and semantic aspects and pulling them closer to the original samples in the latent space, the network is able to focus on the principal information of actions.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2023)
Article
Computer Science, Artificial Intelligence
Sravani Yenduri, Nazil Perveen, Vishnu Chalavadi, Krishna C. Mohan
Summary: The study presents a dynamic kernel-based approach for effective recognition of fine-grained actions by extracting local spatio-temporal features and analyzing them using a Gaussian mixture model. Kernels are built to find the similarity between fine-grained actions by mapping the statistics to the kernel feature space, and the effectiveness of the proposed method is demonstrated through experiments.
PATTERN RECOGNITION
(2022)
Article
Robotics
Minseok Seo, Donghyeon Cho, Sangwoo Lee, Jongchan Park, Daehan Kim, Jaemin Lee, Jingi Ju, Hyeoncheol Noh, Dong-Geol Choi
Summary: The common paradigm of CNN-based action recognition models to use average of dense predictions is inefficient. We propose a lightweight CNN-based sampler that selects relevant frames for action recognition, resulting in improved performance and reduced computation.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2022)
Article
Computer Science, Artificial Intelligence
Peiqin Zhuang, Yu Guo, Zhipeng Yu, Luping Zhou, Lei Bai, Ding Liang, Zhiyong Wang, Yali Wang, Wanli Ouyang
Summary: Motion modeling plays a crucial role in modern action recognition methods. However, variations in motion dynamics across different video clips present a challenge in adaptively covering proper motion information. In this paper, we propose a Motion Diversification and Selection (MoDS) module that generates diversified spatio-temporal motion features and dynamically selects the suitable motion representation for categorizing input videos. Our method achieves state-of-the-art performance on benchmarks with large motion variations.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2022)
Article
Allergy
Changyi Xu, Lijuan Du, Yubiao Guo, Yuxia Liang
Summary: This study aimed to identify differentially expressed circRNAs in induced sputum cells of asthma patients, in order to provide potential biomarkers and insights for asthma research. Through high-throughput sequencing, the differentially expressed circRNAs in asthma patients were screened, and their expression was verified by qRT-PCR. The correlation between circRNA and asthma was analyzed, and a possible ceRNA network was predicted and analyzed.
INTERNATIONAL ARCHIVES OF ALLERGY AND IMMUNOLOGY
(2023)
Article
Engineering, Electrical & Electronic
Fei Li, Jiangbin Zheng, Yuan-fang Zhang, Wenjing Jia, Qianru Wei, Xiangjian He
Summary: The poor quality of underwater images affects underwater development projects. Deep learning-based techniques are successful in image restoration and enhancement, but the limited availability of paired training data and vivid color correction pose challenges. This study proposes an unsupervised training strategy, using a GAN-based framework, to generate high-quality images without paired data. Experimental results show superior performance in underwater benchmarks.
SIGNAL PROCESSING-IMAGE COMMUNICATION
(2023)
Article
Computer Science, Interdisciplinary Applications
Chengpei Xu, Wenjing Jia, Ruomei Wang, Xiangjian He, Baoquan Zhao, Yuanfang Zhang
Summary: With the lack of elaborating annotations and interesting content in educational videos, this article proposes a slide-based video navigation tool that extracts the hierarchical structure and semantic relationship of visual entities in videos by integrating multichannel information. Through a novel deep learning framework, features of visual entities are extracted from presentation slides, and a clustering approach is used to determine the hierarchical relationships between these entities. By evaluating their semantic relationships, visual entities are associated with their corresponding audio speech text, generating a multilevel table of contents and notes for improved learning navigation.
IEEE TRANSACTIONS ON LEARNING TECHNOLOGIES
(2023)
Article
Environmental Sciences
Yue Xi, Wenjing Jia, Qiguang Miao, Junmei Feng, Xiangzeng Liu, Fei Li
Summary: Benefiting from the advances in object detection in remote sensing, this study proposes a Collaborative Deraining Network (CoDerainNet) that simultaneously trains a deraining subnetwork and a droneDet subnetwork to improve the accuracy of object detection in rainy weather conditions (Rainy DroneDet). Additionally, a Collaborative Teaching paradigm (ColTeaching) is introduced to remove rain-specific interference and improve detection performance. Experimental results show that CoDerainNet can reduce computational costs while maintaining comparable detection performance to state-of-the-art models.
Article
Computer Science, Information Systems
Muhammad Babar, Mian Ahmad Jan, Xiangjian He, Muhammad Usman Tariq, Spyridon Mastorakis, Ryan Alturki
Summary: With the rise of IoT, the awareness of edge computing is gaining importance. However, edge computing faces challenges in tackling the diverse applications of IoT due to the massive heterogeneous data they produce. To address these challenges, we propose an optimized IoT-enabled big data analytics architecture for edge-cloud computing using machine learning.
IEEE INTERNET OF THINGS JOURNAL
(2023)
Article
Computer Science, Artificial Intelligence
Jiachen Kang, Wenjing Jia, Xiangjian He
Summary: Existing deep learning models often have performance drop when facing out-of-distribution tasks in computer vision. This study presents a new training methodology using synthetic datasets to enable deep neural networks to acquire generalizable knowledge. The learned knowledge is effectively utilized through the proposed InterpretNet architecture, resulting in a significant improvement in image classification accuracy.
Article
Computer Science, Information Systems
Chengpei Xu, Wenjing Jia, Tingcheng Cui, Ruomei Wang, Yuan-fang Zhang, Xiangjian He
Summary: This paper improves classic bottom-up text detection frameworks by fusing visual-relational features, developing effective false positive/negative suppression mechanisms, and introducing a new shape-approximation strategy, resulting in enhanced performance compared to state-of-the-art methods.
IEEE TRANSACTIONS ON MULTIMEDIA
(2023)
Article
Computer Science, Information Systems
Chengpei Xu, Wenjing Jia, Ruomei Wang, Xiaonan Luo, Xiangjian He
Summary: Bottom-up text detection methods play an important role in arbitrary-shape scene text detection. However, this paper proposes a novel approach named MorphText to capture the regularity of texts using deep morphology. By designing two deep morphological modules, text segments can be regularized and reliable connections can be determined. Experimental results show that MorphText outperforms existing methods on multiple benchmark datasets.
IEEE TRANSACTIONS ON MULTIMEDIA
(2023)
Proceedings Paper
Computer Science, Information Systems
Han Xu, Priyadarsi Nanda, Jie Liang, Xiangjian He
Summary: This paper analyzes how dual ethical risk affects the performance of federated learning schemes and proposes a solution, an optimal multi-stage contract-theoretic incentive mechanism, to mitigate the risk. This is the first discussion on the dual ethical risk for federated learning participants and the first development of an optimal incentive mechanism to address this issue.
NETWORK AND SYSTEM SECURITY, NSS 2022
(2022)
Proceedings Paper
Computer Science, Artificial Intelligence
Nazar Waheed, Muhammad Ikram, Saad Sajid Hashmi, Xiangjian He, Priyadarsi Nanda
Summary: Web-based chatbots provide website owners with increased sales and customer insight, but they also pose risks to user privacy and security. A large-scale analysis reveals that Intercom and LiveChat are the most commonly used chatbots, with differences in the number of third-party domains embedded. Some chatbots use insecure protocols and rely heavily on cookies for tracking and advertisement purposes.
WEB INFORMATION SYSTEMS ENGINEERING - WISE 2022
(2022)
Proceedings Paper
Computer Science, Artificial Intelligence
Ye Huang, Di Kang, Liang Chen, Xuefei Zhe, Wenjing Jia, Linchao Bao, Xiangjian He
Summary: This paper proposes a universal Class-Aware Regularization (CAR) approach to optimize intra-class variance and inter-class distance for more effective utilization of class-level information, resulting in improved accuracy of existing network modules. Experimental results on multiple benchmark datasets demonstrate the superior generalization ability of the proposed method.
COMPUTER VISION - ECCV 2022, PT XXVIII
(2022)