Article
Computer Science, Information Systems
Yun-Hao Yuan, Jin Li, Yun Li, Jipeng Qiang, Bin Li, Wankou Yang, Furong Peng
Summary: The proposed OPLS-SR method is an effective FSR learning approach based on orthonormalized partial least squares, which learns a latent coherent feature space of low-dimensional HR and LR face embeddings and super-resolves LR face images through global face reconstruction and facial detail compensation. Experimental results show the effectiveness of the method on face databases in terms of quantitative and qualitative evaluations.
INFORMATION SCIENCES
(2021)
Article
Computer Science, Artificial Intelligence
Xiaoguang Li, Ning Dong, Jianglu Huang, Li Zhuo, Jiafeng Li
Summary: A discriminative self-attention cycle generative adversarial network is proposed for real-world face image super-resolution, utilizing unpaired samples to train the degradation network and reconstruction network simultaneously, while introducing self-attention mechanism and Siamese face recognition network for better performance.
IET IMAGE PROCESSING
(2021)
Article
Computer Science, Artificial Intelligence
Muhammad Farooq, Matthew N. Dailey, Arif Mahmood, Jednipat Moonrinta, Mongkol Ekpanyapong
Summary: This paper proposes a novel approach to super-resolution by utilizing HR and LR datasets sharing important properties, and acquiring roughly paired training data from the main stream and sub-stream of the same IP camera. Utilizing cycle generative adversarial networks (Cycle GANs), impressive super-resolved images for low-quality test images are generated, enabling the extraction of high-quality face images from low-resolution video streams. Developers of various applications such as face recognition, 3D face reconstruction, face alignment, and more will benefit from the methods introduced in this work.
NEURAL COMPUTING & APPLICATIONS
(2021)
Article
Computer Science, Hardware & Architecture
Banti Kumar, Shyam Singh Rajput
Summary: A new low-light robust face super-resolution (SR) model is proposed to improve the robustness of current SR algorithms against low-light problems. It utilizes dictionary images to detect the low-light effects and applies the same amount of effect to low-resolution dictionary images for minimizing the reconstruction error. The introduction of morphological transformation further improves the reconstruction weights.
COMPUTERS & ELECTRICAL ENGINEERING
(2023)
Article
Computer Science, Theory & Methods
Junjun Jiang, Chenyang Wang, Xianming Liu, Jiayi Ma
Summary: This survey systematically reviews deep learning-based face super-resolution (FSR) methods. It summarizes the problem formulation of FSR, introduces assessment metrics and loss functions. It elaborates on facial characteristics and popular datasets used in FSR, and categorizes existing methods based on the utilization of facial characteristics. For each category, it provides a general description of design principles, an overview of representative approaches, and discusses their pros and cons. The survey also evaluates the performance of state-of-the-art methods and introduces joint FSR and other tasks, as well as FSR-related applications, while envisioning future technological advancements in this field.
ACM COMPUTING SURVEYS
(2023)
Article
Computer Science, Artificial Intelligence
Jonghyun Kim, Gen Li, Inyong Yun, Cheolkon Jung, Joongkyu Kim
Summary: This paper proposes a novel Edge and Identity Preserving Network for Face SR Network, EIPNet, which aims to minimize distortion in facial components and generate high-quality 8x scaled SR images. By utilizing an edge block, identity loss function, and luminance-chrominance error method, the network is able to achieve this goal without requiring extra labels, long training time, and large computation memory. Extensive experiments demonstrate that the proposed network outperforms state-of-the-art methods on challenging datasets.
Article
Computer Science, Information Systems
Shengxiang Luo, Jinbo Lu
Summary: This paper introduces a face super-resolution network based on gradient information compensation, which can generate high-resolution face images from low-resolution images. The network compensates gradient features using feature residual blocks and gradient extraction blocks, and introduces a feature fusion mechanism to improve performance.
Article
Computer Science, Artificial Intelligence
Kangli Zeng, Zhongyuan Wang, Tao Lu, Jianyu Chen, Jiaming Wang, Zixiang Xiong
Summary: This paper proposes a self-attention learning network (SLNet) for three-stage face super-resolution, which fully explores the interdependence of low- and high-level spaces to achieve compensation of the information used for reconstruction. The experimental results show that SLNet has a competitive performance compared to the state-of-the-art methods.
Article
Computer Science, Information Systems
Shyam Singh Rajput
Summary: An intelligent surveillance system faces challenges in processing noisy low-resolution images, leading to the development of face super-resolution techniques. The proposed algorithm effectively suppresses noise and preserves key features of input faces, showing better reconstruction capability compared to state-of-the-art models.
MULTIMEDIA TOOLS AND APPLICATIONS
(2022)
Article
Computer Science, Artificial Intelligence
Anurag Singh Tomar, K. V. Arya, Shyam Singh Rajput
Summary: This paper proposes a new noise robust face super-resolution model using an attentive ExFeat-based generative adversarial network. The model introduces the Exigent Feature Attention Unit (ExFAU) consisting of an Exigent Feature (ExFeat) block with a spatial attention unit to enhance the visual quality of the generated face images. Experimental outcomes show that the proposed model achieves state-of-the-art performance on standard datasets.
PATTERN RECOGNITION LETTERS
(2023)
Article
Engineering, Electrical & Electronic
Chenyang Wang, Junjun Jiang, Zhiwei Zhong, Xianming Liu
Summary: This method proposes a novel approach to learn facial prior knowledge during the training stage and only use low-resolution face images during testing, effectively overcoming the difficulty of prior estimation. By directly exploring high-quality facial prior and progressively propagating this knowledge, the method outperforms existing face super-resolution methods on benchmark datasets.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2022)
Article
Computer Science, Information Systems
Surendra Nagar, Ankush Jain, Pramod Kumar Singh, Ajay Kumar
Summary: This paper proposes a novel noise-robust face super-resolution method that estimates the true noise level and uses a deep convolutional neural network to process low-resolution noisy images. The experimental results show that the proposed method outperforms existing methods.
MULTIMEDIA TOOLS AND APPLICATIONS
(2023)
Article
Computer Science, Information Systems
Ji-Soo Kim, Keunsoo Ko, Chang-Su Kim
Summary: In this paper, a novel face super-resolution network called CollageNet is proposed for generating high-resolution facial images by utilizing reference images at the patch level, showcasing state-of-the-art performances in experimental results.
Article
Computer Science, Information Systems
Yuantao Chen, Volachith Phonevilay, Jiajun Tao, Xi Chen, Runlong Xia, Qian Zhang, Kai Yang, Jie Xiong, Jingbo Xie
Summary: The method uses deep residual networks and deep neural networks to reconstruct face images in high resolution, enhancing performance through the extraction of multilevel features and improvement of network architecture to ultimately achieve high-resolution face images.
MULTIMEDIA TOOLS AND APPLICATIONS
(2021)
Article
Computer Science, Artificial Intelligence
Huan Wang, Qian Hu, Chengdong Wu, Jianning Chi, Xiaosheng Yu, Hao Wu
Summary: This paper proposes a novel CNN-based Dual Closed loop Network (DCLNet) for face super-resolution, which utilizes dual learning networks to constrain the possible mapping space and improve the reconstruction performance. Experimental results demonstrate that the proposed method achieves state-of-the-art performance on the CelebA and Helen datasets.
KNOWLEDGE-BASED SYSTEMS
(2021)
Article
Geochemistry & Geophysics
Chen Ma, Junjun Jiang, Huayi Li, Wenxue Cui, Guoyuan Li
Summary: In recent years, CNN-based methods have shown limitations in modeling spectral-wise long-range dependencies, while transformer-based deep learning methods have shown superiority in this aspect. However, the special tokenization of transformer-based methods leads to the involvement of redundant tokens, which do not contribute positively to classification. To solve this problem, a lightweight SSPT framework is proposed to efficiently extract spatial-spectral features of HSI by progressively reducing redundant tokens.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Engineering, Electrical & Electronic
Deepak Rai, Shyam Singh Rajput
Summary: Face hallucination (FH) techniques generate high-resolution face images from low-resolution images. Existing FH techniques cannot handle motion blur. This article presents a learning-based FH algorithm called Motion Blur Embedded Nearest Proximate Patch Representation (MBENPPR) to address this issue. The MBENPPR algorithm estimates the motion blur kernel and embeds it in training images to reduce the effect of motion blur. It also selects nearest proximate patches to preserve sharp edges and texture information. Simulation results show that MBENPPR outperforms existing algorithms.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2023)
Article
Biochemistry & Molecular Biology
Xiaopeng Hu, Sanqi An, Jiemei Chu, Bingyu Liang, Yanyan Liao, Junjun Jiang, Yao Lin, Li Ye, Hao Liang
Summary: The monkeypox outbreak is a global public health emergency, and the lack of valid and safe medicine is a crucial obstacle in eradicating orthopoxvirus infections. The identification of potential inhibitors from natural products, including Traditional Chinese Medicine, through molecular modeling, could expand the arsenal of antiviral chemotherapeutic agents.
Article
Automation & Control Systems
Xinya Wang, Jiayi Ma, Junjun Jiang
Summary: Previous deep learning-based super-resolution methods rely on predefined degradation processes and may suffer from deterioration when the real degradation is inconsistent. In this paper, we propose a contrastive regularization method that exploits blurry and clear images as negative and positive samples, respectively, to improve blind super-resolution performance. We also extract global statistical prior information instead of estimating degradation to capture the distortion characteristics and make our method adaptive to changes in distortions. Experimental results demonstrate that our lightweight CRDNet surpasses state-of-the-art blind super-resolution approaches.
IEEE-CAA JOURNAL OF AUTOMATICA SINICA
(2023)
Article
Computer Science, Theory & Methods
Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, Xiangyang Ji
Summary: This survey provides a comprehensive overview of recent progress in guided depth map super-resolution (GDSR). It covers the problem background, challenges, commonly used datasets and evaluation methods. Different categories of methods, including filtering-based, prior-based, and learning-based, are introduced along with representative approaches and their applications. The performance of representative methods is evaluated through experiments conducted with unified configurations, and possible research directions and open problems are discussed.
ACM COMPUTING SURVEYS
(2023)
Article
Engineering, Electrical & Electronic
Yuanzhi Wang, Tao Lu, Yanduo Zhang, Zhongyuan Wang, Junjun Jiang, Zixiang Xiong
Summary: In this paper, the authors propose a method called FaceFormer, which combines global features from Transformers and local features from CNNs to restore high-quality face images.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2023)
Article
Computer Science, Artificial Intelligence
Xingyu Hu, Junjun Jiang, Xianming Liu, Jiayi Ma
Summary: Multi-focus image fusion (MFF) is a challenging task due to the difficulty in distinguishing different blur levels and the lack of real supervised data. In this study, we propose a novel deep learning-based framework named ZMFF, which captures the deep prior of the fused image and the focus map using deep image prior and deep mask prior networks, respectively. Our method achieves promising performance, generalization, and flexibility on both synthetic and real-world datasets without the need for extensive training data.
INFORMATION FUSION
(2023)
Article
Computer Science, Information Systems
Anurag Singh Tomar, K. V. Arya, Shyam Singh Rajput
Summary: This paper proposes a face super-resolution framework based on the feature attention unit. The introduced ExFeat block with spatial attention helps in learning detailed features and reducing noise. Experimental results show that the proposed framework outperforms other competitive methods on CelebAHQ and LFW face datasets.
MULTIMEDIA TOOLS AND APPLICATIONS
(2023)
Article
Computer Science, Artificial Intelligence
Anurag Singh Tomar, K. V. Arya, Shyam Singh Rajput
Summary: This paper proposes a new noise robust face super-resolution model using an attentive ExFeat-based generative adversarial network. The model introduces the Exigent Feature Attention Unit (ExFAU) consisting of an Exigent Feature (ExFeat) block with a spatial attention unit to enhance the visual quality of the generated face images. Experimental outcomes show that the proposed model achieves state-of-the-art performance on standard datasets.
PATTERN RECOGNITION LETTERS
(2023)
Article
Environmental Sciences
Ziqian Liu, Wenbing Wang, Qing Ma, Xianming Liu, Junjun Jiang
Summary: In this paper, a full 3D convolutional neural network (F3DUN) is proposed for hyperspectral image super-resolution (HSISR) tasks. The F3DUN model combined with the U-Net architecture achieves state-of-the-art performance on HSISR tasks by utilizing skip connections and multi-scale features. Additionally, the paper compares F3DUN with a 3D/2D mixed model and finds that the full 3D CNN has a larger capacity and can obtain better results with the same number of parameters. Furthermore, experimental results demonstrate that the full 3D CNN model is less sensitive to data scaling and outperforms the 3D/2D mixed model on small-scale datasets.
Article
Computer Science, Artificial Intelligence
Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Guangcheng Wang, Zhen Han, Junjun Jiang, Zixiang Xiong
Summary: This study focuses on addressing the problem of generating rain-free images under complex rain conditions using deep learning models. By designing a multi-level pyramid structure, non-local fusion module, attention fusion module, and residual learning branch to handle different challenges, the results demonstrate that our method achieves superior performance in generating rain-free images.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Engineering, Electrical & Electronic
Anurag Singh Tomar, K. V. Arya, Shyam Singh Rajput
Summary: This article proposes a novel deep hybrid feature (HyFeat)-based Attention in Attention model for face super-resolution. The proposed model combines a coarse SR network and deep CNN to generate high-resolution images. Experimental results show that the proposed model achieves state-of-the-art performance on standard datasets.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2023)
Proceedings Paper
Computer Science, Artificial Intelligence
Rafael E. Rivadeneira, Angel D. Sappa, Boris X. Vintimilla, Chenyang Wang, Junjun Jiang, Xianming Liu, Zhiwei Zhong, Dai Bin, Li Ruodi, Li Shengye
Summary: This paper presents the results of two tracks from the fourth Thermal Image Super-Resolution (TISR) challenge, including the improvement of the first track compared to last year's challenge, and the use of a new dataset in the second track. The high participation of over 150 teams demonstrates the ongoing interest in this topic in the community.
2023 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS, CVPRW
(2023)
Article
Computer Science, Theory & Methods
Junjun Jiang, Chenyang Wang, Xianming Liu, Jiayi Ma
Summary: This survey systematically reviews deep learning-based face super-resolution (FSR) methods. It summarizes the problem formulation of FSR, introduces assessment metrics and loss functions. It elaborates on facial characteristics and popular datasets used in FSR, and categorizes existing methods based on the utilization of facial characteristics. For each category, it provides a general description of design principles, an overview of representative approaches, and discusses their pros and cons. The survey also evaluates the performance of state-of-the-art methods and introduces joint FSR and other tasks, as well as FSR-related applications, while envisioning future technological advancements in this field.
ACM COMPUTING SURVEYS
(2023)
Article
Engineering, Electrical & Electronic
Alam Abbas Syed, Hassan Foroosh
Summary: This paper presents effective methods using spherical polar Fourier transform data for two different applications: 3D volumetric registration and machine learning classification network. The proposed method for registration offers unique and effective techniques, handling arbitrary large rotation angles and showing robustness. The modified classification network achieves robust classification results in processing spherical data.
Article
Engineering, Electrical & Electronic
Ruibo Fan, Mingli Jing, Jingang Shi, Lan Li, Zizhao Wang
Summary: In this study, a new low-rank sparse decomposition algorithm named TVRPCA+ is proposed for foreground-background separation. The algorithm combines spectral norm, structured sparse norm, and total variation regularization to suppress noise and obtain cleaner foregrounds. Experimental results demonstrate that TVRPCA+ achieves high performance in complex backgrounds and noise scenarios.
Article
Engineering, Electrical & Electronic
Omair Aldimashki, Ahmet Serbes
Summary: This paper proposes a coarse-to-fine FrFT-based algorithm for chirp-rate estimation of multi-component LFM signals, which achieves improved performance and a reduced signal-to-noise breakdown threshold by utilizing mathematical models for coarse estimation and a refined estimate-and-subtract strategy. Extensive simulation results demonstrate that the proposed algorithm performs very close to the Cramer-Rao lower bound, with the advantages of eliminating leakage effect, avoiding error propagation, and maintaining acceptable computational cost compared to other state-of-the-art methods.
Article
Engineering, Electrical & Electronic
Xinlei Shi, Xiaofei Zhang, Yuxin Sun, Yang Qian, Jinke Cao
Summary: In this paper, a low-complexity localization approach for multiple sources using two-dimensional discrete Fourier transform (2D-DFT) is proposed. The method computes the cross-covariance and utilizes phase offset method and total least square solution to obtain accurate position estimates.
Article
Engineering, Electrical & Electronic
Prabhanjan Mannari, Ratnasingham Tharmarasa, Thiagalingam Kirubarajan
Summary: This paper discusses the problem of extended target tracking for a single 2D extended target with a known convex polytope shape and dynamics. It proposes a framework based on the existing point multitarget tracking framework to address the challenges of uncertainty in shape and kinematics, as well as self-occlusion. The algorithm developed using this framework is capable of dynamically changing the number of parameters used to describe the shape and estimating the whole target shape even when different parts of the target are visible at different frames.
Article
Engineering, Electrical & Electronic
Yongsong Li, Zhengzhou Li, Jie Li, Junchao Yang, Abubakar Siddique
Summary: This paper proposes a weighted adaptive ring top-hat transformation (WARTH) for extracting infrared small targets in complex backgrounds. The WARTH method effectively measures local and global feature information using an adaptive ring-shaped structural element and a target awareness indicator, resulting in accurate detection of small targets with minimized false alarms.
Article
Engineering, Electrical & Electronic
Yu Wang, Zhen Qin, Jun Tao, Yili Xia
Summary: In this paper, an enhanced sparsity-aware recursive least squares (RLS) algorithm is proposed, which combines the proportionate updating (PU) and zero-attracting (ZA) mechanisms, and introduces a general convex regularization (CR) function and variable step-size (VSS) technique to improve performance.
Article
Engineering, Electrical & Electronic
Neil J. Bershad, Jose C. M. Bermudez
Summary: This paper analyzes the impact of processing delay on the Least Mean Squares (LMS) algorithm in system identification, highlighting bias issues in the resulting weight vector.
Article
Engineering, Electrical & Electronic
Kanghui Jiang, Defu Jiang, Mingxing Fu, Yan Han, Song Wang, Chao Zhang, Jingyu Shi
Summary: In this paper, a novel method for velocity estimation using multicarrier signals in a single dwell is proposed, which effectively addresses the issue of Doppler ambiguity in pulse Doppler radars.
Article
Engineering, Electrical & Electronic
Xiao-Jun Zhang, Peng-Lang Shui, Yu-Fan Xue
Summary: This paper proposes a method for low-velocity small target detection in maritime surveillance radars. It models sea clutter sequences using the spherical invariant random vector (SIRV) model with block tridiagonal speckle covariance matrix and inverse Gamma distributed texture. The proposed detector, which is a long-time adaptive generalized likelihood ratio test with linear threshold detector (GLRT-LTD), shows competitive detection performance in experiments.
Article
Engineering, Electrical & Electronic
Aiyi Zhang, Fulai Liu, Ruiyan Du
Summary: This paper proposes an adaptive weighted robust data recovery method with total variation regularization for hyperspectral image. The method models the HSI recovery problem as a tensor robust principal component analysis optimization problem, decomposing the data into low-rank HSI data, outliers, and noise component. An adaptive weighted strategy is then defined to impose on the tensor nuclear norm and outliers, using the priori information of singular values and strengthening the sparsity of outliers.
Article
Engineering, Electrical & Electronic
Hamid Asadi, Babak Seyfe
Summary: This paper presents a novel approach for estimating the model order in the presence of observation errors. The proposed method is based on correntropy estimation of eigenvalues in the observation space, which is further enhanced by resampling the observations using the bootstrap method. The algorithm partitions the observation space into signal and noise subspaces using the covariance matrix of mixtures, and determines the model order based on a correntropy estimator with kernel functions. Theoretical analysis and comparative evaluations demonstrate the superiority of this information-theoretic approach.
Article
Engineering, Electrical & Electronic
Buket colak Guvenc, Engin Cemal Menguc
Summary: In this paper, a novel family of online censoring based complex-valued least mean kurtosis (CLMK) algorithms is proposed. The algorithms censor less informative complex-valued data streams and reduce the costs of data processing without affecting accuracy. Robust algorithms are also developed to handle outliers. The simulation results confirm the attractive features of the proposed algorithms in large-scale system identification and regression scenarios.
Article
Engineering, Electrical & Electronic
Yun Su, Weixian Tan, Yifan Dong, Wei Xu, Pingping Huang, Jianxin Zhang, Diankun Zhang
Summary: In this study, a novel method for detecting low-resolution and small targets in millimeter wave radar images is proposed. The Wavelet-Conv structure and Wavelet-Attention mechanism are introduced to overcome the limitations of existing detectors. Experimental results demonstrate that the proposed method improves recall and mean average precision while maintaining competitive inference speed.
Article
Engineering, Electrical & Electronic
Xin Wang, Xingxing Jiang, Qiuyu Song, Jie Liu, Jianfeng Guo, Zhongkui Zhu
Summary: This study proposes a variational mode extraction (VME) method for extracting specific modes from complicated signals. By exploring the convergence property of VME, strategies for identifying ICF and determining the balance parameter are designed, and a bandwidth estimation strategy is constructed. The effectiveness of the proposed method for bearings fault diagnosis is verified and compared with other methods.