Article
Computer Science, Artificial Intelligence
Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, Junchi Yan
Summary: Image matching is a fundamental task in various visual applications, and with the development of deep learning techniques, there has been an increasing number of methods proposed in this field. However, the challenge remains in choosing the suitable method for specific applications and designing image matching methods with superior performance. This comprehensive review and analysis provide insights into classical and latest techniques, and offer prospects for future development in image matching technologies.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2021)
Article
Engineering, Electrical & Electronic
Xianfeng Song, Yi Zou, Zheng Shi, Yanfeng Yang
Summary: With the popularity of vision sensors and advancement in image processing technology, machine vision tasks based on image matching have gained significant attention. This study proposes a novel image matching scheme that combines handcrafted and deep features using a fine-grained decision-level fusion (DLF) method. Experimental results show significant improvements in correct matches and robustness, and the versatility of the proposed scheme is demonstrated through a target localization algorithm incorporating an unmanned aerial vehicle (UAV).
IEEE SENSORS JOURNAL
(2023)
Article
Computer Science, Artificial Intelligence
Jiale Cao, Yanwei Pang, Jin Xie, Fahad Shahbaz Khan, Ling Shao
Summary: Pedestrian detection is a challenging problem in computer vision, and this survey comprehensively reviews recent advances in this field. It covers single-spectral and multi-spectral pedestrian detection methods, as well as the use of handcrafted and deep features. The survey also introduces relevant datasets and evaluation metrics, and provides deep experimental analysis. Open problems and future research directions are highlighted in the conclusion.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
(2022)
Article
Biochemical Research Methods
Cesar R. Garcia-Jacas, Luis A. Garcia-Gonzalez, Felix Martinez-Rios, Issac P. Tapia-Contreras, Carlos A. Brizuela
Summary: Research shows that non-handcrafted features outperform handcrafted features in terms of performance, but a performance improvement is achieved when both types of features are merged. Non-handcrafted features have higher information content, while handcrafted features are more important, indicating complementarity between the two types of features.
BRIEFINGS IN BIOINFORMATICS
(2022)
Article
Computer Science, Artificial Intelligence
Victor Hugo da Silva Muniz, Joao Baptista de Oliveira e Souza Filho
Summary: This paper discusses the importance of music genre in music recommendations and presents a method to improve system performance through the generation of new handcrafted features and feature selection.
NEURAL COMPUTING & APPLICATIONS
(2023)
Article
Computer Science, Information Systems
Hassan Shahmohammadi, MirHossein Dezfoulian, Muharram Mansoorizadeh
Summary: The article presents a deep learning-based model for paraphrase detection, which encodes input data and measures sentence similarity using handcrafted features, achieving good evaluation results.
MULTIMEDIA TOOLS AND APPLICATIONS
(2021)
Article
Mathematics, Interdisciplinary Applications
Carlos Frederico S. da F. Mendes, Renato A. Krohling
Summary: Skin lesions diagnostic is a challenging problem, and our study shows that combining features from convolutional neural networks, handcrafted features and clinical information can improve the automated diagnosis of skin cancer. Using a clinical image dataset with patient information, we demonstrate that clinical features as a complement to CNN and handcrafted features significantly enhance the accuracy in diagnosing cancer and melanoma.
CHAOS SOLITONS & FRACTALS
(2022)
Article
Computer Science, Information Systems
Lachlan Burne, Chiranjibi Sitaula, Archana Priyadarshi, Mark Tracy, Omid Kavehei, Murray Hinder, Anusha Withana, Alistair McEwan, Faezeh Marzbanrad
Summary: This article proposes a novel technique for automated peristalsis sound detection from neonatal abdominal sound recordings and compares it to various other machine learning approaches. It adopts an ensemble approach that utilizes handcrafted as well as one and two dimensional deep features obtained from Mel Frequency Cepstral Coefficients (MFCCs). The results show that our method provides an accuracy of 95.1% and an Area Under Curve (AUC) of 85.6%, outperforming both the baselines and the recent works significantly. These encouraging results demonstrate that our proposed Ensemble-based Deep Learning model is helpful for neonatologists to facilitate tele-health applications.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
(2023)
Article
Computer Science, Artificial Intelligence
Guang-Hai Liu, Zuo-Yong Li, Jing-Yu Yang, David Zhang
Summary: This article introduces a novel image retrieval method that improves retrieval performance by using sublimated deep features. The method incorporates orientation-selective features and color perceptual features, effectively mimicking these mechanisms to provide a more discriminating representation.
PATTERN RECOGNITION
(2024)
Article
Biology
Betul Ay, Cihan Turker, Elif Emre, Kevser Ay, Galip Aydin
Summary: This paper introduces the characteristics and harm of nasal polyps and proposes a reliable rhinology assistance system for recognizing them. The authors design a new dataset including 80 participants and conduct experiments using machine learning and deep learning algorithms. They find that deep learning algorithms achieve high accuracy in identifying nasal polyps. The research results are significant for supporting clinical decision systems.
COMPUTERS IN BIOLOGY AND MEDICINE
(2022)
Article
Computer Science, Information Systems
Misaj Sharafudeen, Vinod S. S. Chandra
Summary: This article proposes a skin cancer diagnosis system that combines conventional therapeutic methods and deep learning frameworks. The system utilizes image data, handcrafted features, and patient metadata to effectively diagnose skin cancers. The method of ensemble learning and weighting models based on their contribution is used to improve accuracy and sensitivity.
MULTIMEDIA TOOLS AND APPLICATIONS
(2023)
Article
Computer Science, Artificial Intelligence
Feiwei Qin, Shi Qiu, Shuming Gao, Jing Bai
Summary: A 3D CAD model retrieval approach that combines sketches and unsupervised learning is proposed in this paper, utilizing automatic structural semantics capture algorithms and deep variational autoencoders for fast, accurate, and easy CAD model matching. Experiments validate the feasibility and effectiveness of the proposed approach.
ADVANCED ENGINEERING INFORMATICS
(2022)
Article
Multidisciplinary Sciences
Maroua Tounsi, Ikram Moalla, Umapada Pal, Adel M. Alimi
Summary: The study compares the effectiveness of handcrafted and hybrid features in recognizing Latin/Arabic text in natural scenes. By combining deep-learned features with handcrafted features, the recognition accuracy is significantly improved.
ARABIAN JOURNAL FOR SCIENCE AND ENGINEERING
(2022)
Article
Chemistry, Analytical
Hongzu Li, Pierre Boulanger
Summary: Cardiovascular diseases are a leading cause of death worldwide, and early detection and treatment are crucial. This paper proposes a new method that combines a Short-Time Fourier Transform (STFT) spectrogram with handcrafted features to detect heart anomalies beyond the capabilities of commercial products.
Article
Computer Science, Information Systems
Savita Walia, Krishan Kumar, Munish Kumar, Xiao-Zhi Gao
Summary: Content authentication of digital images has attracted attention due to the increase in multimedia data dissemination through the Internet. Various methods, including feature engineering and deep learning, have been explored for image forgery detection, with deep learning achieving up to 98% accuracy. A fusion-based approach combining handcrafted and deep features obtained a 99.3% accuracy on benchmark datasets.
Article
Computer Science, Artificial Intelligence
Bo Ni, Zhiyuan Liu, Xiantao Cai, Michele Nappi, Shaohua Wan
Summary: This paper proposes a novel deformable contour model for segmenting ultrasound image sequences. The model utilizes the power of deep learning network in learning image features to overcome the challenges in ultrasound image segmentation. Experimental results show that the proposed method outperforms state-of-the-art methods in clinical ultrasound images.
NEURAL COMPUTING & APPLICATIONS
(2023)
Article
Computer Science, Information Systems
Yue Zhang, Fanghui Zhang, Yi Jin, Yigang Cen, Viacheslav Voronin, Shaohua Wan
Summary: In this paper, a novel local correlation ensemble model is proposed to address the cross-domain problem in the Re-ID task. The model improves the utilization of unlabeled samples in the target domain by focusing on person's features and calculating the distance between nodes. Experimental results on large-scale public Re-ID datasets demonstrate the effectiveness of the proposed method.
ACM TRANSACTIONS ON MULTIMEDIA COMPUTING COMMUNICATIONS AND APPLICATIONS
(2023)
Article
Computer Science, Artificial Intelligence
Yirui Wu, Hao Li, Lilai Zhang, Chen Dong, Qian Huang, Shaohua Wan
Summary: Accurately understanding low-resource languages is crucial for task-oriented human-computer dialogue systems. This involves intent detection and slot filling, which face challenges due to semantic ambiguity and implicit intentions. To address these issues, a joint intent detection method using asynchronous training strategy is proposed, which encodes local text information and emphasizes relationship among words. By fusing hidden states or fine-tuning the network with key information, the relevance between intent detection and slot filling is greatly improved. Validation on airline travel (ATIS) and electricity service (ECSF) datasets achieves 97.49% and 89.68% accuracy, respectively, confirming the effectiveness of joint learning and asynchronous training.
ACM TRANSACTIONS ON ASIAN AND LOW-RESOURCE LANGUAGE INFORMATION PROCESSING
(2023)
Article
Engineering, Civil
Chen Chen, Chenyu Wang, Bin Liu, Ci He, Li Cong, Shaohua Wan
Summary: Edge intelligence technology combined with computer vision can improve traffic information processing and enhance vehicle detection ability. Additionally, using an improved image segmentation algorithm helps reduce network size while improving segmentation accuracy.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2023)
Article
Computer Science, Information Systems
Qian Cheng, Yirui Wu, Aniello Castiglione, Fabio Narducci, Shaohua Wan
Summary: This paper introduces a deep learning-based flood prediction method, presenting a dual attention embedding network (DA-Net). The proposed method utilizes a convolution self-attention module (CSA) and a Temporal-related Feature Attention (TFA) module to capture both local and global flood features, achieving accurate prediction results.
JOURNAL OF SIGNAL PROCESSING SYSTEMS FOR SIGNAL IMAGE AND VIDEO TECHNOLOGY
(2023)
Article
Computer Science, Artificial Intelligence
Yirui Wu, Hao Li, Xi Feng, Andrea Casanova, Andrea F. Abate, Shaohua Wan
Summary: Deep learning methods have shown significant performance in medical image analysis tasks, but lack interpretability in feature extraction and decision processes. To address this, a novel Group-Disentangled Representation Learning framework (GDRL) is proposed, which disentangles latent space into disease concepts with abundant and non-overlapping feature explanations. By emphasizing the linking relationship between semantical concepts of disease and low-level visual features, GDRL enhances interpretability and showcases potential in predicting diseases from chest X-ray images.
PATTERN RECOGNITION LETTERS
(2023)
Article
Chemistry, Analytical
Aikaterini I. Griva, Achilles D. Boursianis, Shaohua Wan, Panagiotis Sarigiannidis, Konstantinos E. Psannis, George Karagiannidis, Sotirios K. Goudos
Summary: The implementation of smart networks has been greatly advanced by the development of IoT, with LoRa being a prominent technology due to its long-distance transmission capabilities with low power consumption. This study simulated various environments to assess network performance based on different factors and parameters. Path loss model, deployment area size, transmission power, spreading factor, number of nodes and gateways, and antenna gain significantly affect the energy consumption and data extraction rate of LoRa networks. The research performed simulations using the FLoRa framework in OMNeT++, investigating rural and urban environments, as well as a parking area model. The results emphasize the importance of optimizing key parameters for the deployment of smart networks.
Article
Computer Science, Hardware & Architecture
Yirui Wu, Lilai Zhang, Zonghua Gu, Hu Lu, Shaohua Wan
Summary: This article proposes an Edge-AI-driven framework for Facial Expression Recognition (FER) in the wild, addressing challenges such as occlusions, illumination, scale, and head pose variations. It introduces two attention modules, Arbitrary-oriented Spatial Pooling (ASP) and Scalable Frequency Pooling (SFP), for effective feature extraction to improve classification accuracy. The article also presents an edge-cloud joint inference architecture for FER, achieving low-latency inference with a lightweight backbone network on the edge device and optional attention modules partially offloaded to the cloud. Performance evaluation shows a good balance between classification accuracy and inference latency in this approach.
ACM TRANSACTIONS ON EMBEDDED COMPUTING SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Yuling Xi, Ning Wang, Shaohua Wan, Xiaoming Wang, Peng Wang, Yanning Zhang
Summary: Instance segmentation is a visual task that requires predicting per-pixel masks and category labels for each instance. We propose using Neural Architecture Search (NAS) to automatically search for a hardware and memory-friendly feature sharing branch, and our method can be applied to similar multi-task networks. Experimental results show that our method exceeds classical parallel decoder networks in terms of bounding box mAP and segmentation mAP.
APPLIED INTELLIGENCE
(2023)
Article
Computer Science, Information Systems
Rui Zhang, Xiaojie Li, Yifan Zhuo, Kezhong Liu, Xian Zhong, Shaohua Wan
Summary: The Internet of Things (IoT) plays a crucial role in maritime transportation by enabling the construction of marine traffic scenarios, improving efficiency and safety. To address challenges in marine traffic monitoring, a multi-model learning approach is proposed, along with an innovative dataset and a text generation model based on a multi-modal Transformer architecture. Experimental results show that our approach effectively generates accurate and informative descriptions of maritime activity.
COMPUTER COMMUNICATIONS
(2023)
Article
Computer Science, Information Systems
Songtao Ding, Hongyu Wang, Hu Lu, Michele Nappi, Shaohua Wan
Summary: In this paper, a two-path gland segmentation algorithm of colon pathological image based on local semantic guidance is proposed. The improved candidate region search algorithm is employed to generate sub-datasets sensitive to specific features. The semantic feature-guided model is used to extract local adenocarcinoma features and enhance the network's learning ability to gland morphological features.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
(2023)
Editorial Material
Computer Science, Artificial Intelligence
Chinmay Chakraborty, Shaohua Wan, Mohammad R. Khosravi
Summary: Data-driven ontology-based knowledge (OK) presentation and computational linguistics for evolving semantic Asian social networks (ASNs) can provide a robust and real-time data mapping platform, named OK-ASN, that allows massive access across heterogeneous big data sources on the web. It utilizes computational intelligence, web-of-things (WoT) architecture, semantic features, statistical learning and pattern recognition, database management, computer vision, cyber-security, and language processing. OK-ASN is a critical strategy for mining WoT big data and promoting enterprises in various sectors from social media to medical and industrial fields.
ACM TRANSACTIONS ON ASIAN AND LOW-RESOURCE LANGUAGE INFORMATION PROCESSING
(2023)
Article
Automation & Control Systems
Wei Li, Chengchun Gu, Jinlin Chen, Chao Ma, Xiaowu Zhang, Bin Chen, Shaohua Wan
Summary: This paper proposes a data augmentation model called "DLS-GAN" to address the problem of defect location sensitive data augmentation. The model modifies the generator and introduces discriminators, and the experimental results show that DLS-GAN can synthesize high-quality images with desired defects better than state-of-the-art generative models on different types of DLS datasets.
IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING
(2023)
Article
Computer Science, Hardware & Architecture
Chen Chen, Jingfeng Zhang, Xilie Xu, Lingjuan Lyu, Chaochao Chen, Tianlei Hu, Gang Chen
Summary: Adversarial training (AT) is a method to improve the robustness of deep neural networks by training on adversarial variants generated from natural examples. However, as training progresses, the training data becomes less attackable, undermining the enhancement of model robustness. To address this issue, this paper proposes a Decision boundary-aware data Augmentation framework (CODA) that utilizes meta information from previous epochs to guide the augmentation process and generate attackable data close to the decision boundary. CODA outperforms vanilla mixup by providing a higher ratio of attackable data, enhancing model robustness while mitigating the linear behavior between classes that is unfavorable for adversarial training. Experimental results demonstrate that CODA improves adversarial robustness across various training methods and datasets.
IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING
(2023)
Article
Computer Science, Information Systems
Sotirios K. Goudos, Panagiotis D. Diamantoulakis, Achilles D. Boursianis, Panagiotis Sarigiannidis, Konstantinos E. Psannis, Mohammad Abdul Matin, Shaohua Wan, George K. Karagiannidis
Summary: In this work, we address the problem of joint power allocation and user association for non-orthogonal multiple access (NOMA) in downlink networks based on quality-of-service. Due to its non-convex form and the large number of optimization variables, the problem is challenging and we propose two nature-inspired algorithms with low complexity for solving it. We investigate the impact of different network parameters on increasing users and show that evolutionary algorithms are effective in solving this problem, outperforming randomly generated solutions. Furthermore, the advantages of NOMA over OMA become more evident as the number of users increases.