Article
Engineering, Electrical & Electronic
Vineeta Das, Samarendra Dandapat, Prabin Kumar Bora
Summary: The proposed Deep Multi-scale Fusion Convolutional Neural Network (DMF-CNN) effectively encodes multi-scaled disease characteristics by using multiple CNNs with different receptive fields and fusing them, resulting in reliable classification for retinal diseases. The method achieves state-of-the-art performance on publicly available OCT databases and offers an impressive overall accuracy for diagnostic purposes.
IEEE SENSORS JOURNAL
(2021)
Article
Computer Science, Artificial Intelligence
Yingbin Wang, Guanghui Zhao, Kai Xiong, Guangming Shi, Yumeng Zhang
Summary: The utilization of dilated convolution to capture temporal dependencies improves the performance and efficiency of sound event detection systems. The SS-FCN and MS-FCN methods based on fully convolutional networks show competitive performance on different datasets.
Article
Chemistry, Analytical
Changjie Cai, Tomoki Nishimura, Jooyeon Hwang, Xiao-Ming Hu, Akio Kuroda
Summary: This study aimed to accurately detect asbestos using the YOLOv4 model, achieving exceptional performance with a mean average precision of 96.1% +/- 0.4%. Compared to previous software, YOLOv4 demonstrated higher accuracy, precision, recall, and F-1 score, particularly excelling in detecting low fiber concentration samples.
Article
Environmental Sciences
Linbo Tang, Wei Tang, Xin Qu, Yuqi Han, Wenzheng Wang, Baojun Zhao
Summary: A scale-aware feature pyramid network (SARFNet) is proposed for multi-scale object detection within Synthetic Aperture Radar (SAR) images. The network incorporates a feature alignment module, a scale-equalizing pyramid convolution, and a self-learning anchor assignment strategy to flexibly match targets with different appearance changes.
Article
Computer Science, Interdisciplinary Applications
Hans Pinckaers, Wouter Bulten, Jeroen van der Laak, Geert Litjens
Summary: Prostate cancer is the most prevalent cancer among men in Western countries, and pathologists' evaluation is the gold standard for diagnosis. State-of-the-art convolutional neural networks are often patch-based and require detailed pixel-level annotations for effective training.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2021)
Article
Computer Science, Artificial Intelligence
Lei Bi, Jinman Kim, Tingwei Su, Michael Fulham, David Dagan Feng, Guang Ning
Summary: In this study, a deep multi-scale resemblance network (DMRN) was developed for the classification of adrenal masses in CT images. By leveraging paired convolutional neural networks (CNNs) to evaluate intra-class similarities, inter-class separability was improved and the influence of imbalanced training data was reduced through the augmentation of training data. The results demonstrated that the DMRN method outperformed the state-of-the-art approaches in the classification of adrenal masses.
ARTIFICIAL INTELLIGENCE IN MEDICINE
(2022)
Article
Plant Sciences
Yange Sun, Fei Wu, Huaping Guo, Ran Li, Jianfeng Yao, Jianbo Shen
Summary: This paper introduces a novel method called TeaDiseaseNet for tea disease detection. It utilizes a multi-scale self-attention mechanism and a channel attention mechanism to achieve accurate detection and localization of tea disease information. Experimental results demonstrate its superior performance in scenarios with complex backgrounds and varying disease scales, highlighting its potential for intelligent tea disease diagnosis.
FRONTIERS IN PLANT SCIENCE
(2023)
Article
Computer Science, Artificial Intelligence
Jun Zhang, Zhiyuan Hua, Kezhou Yan, Kuan Tian, Jianhua Yao, Eryun Liu, Mingxia Liu, Xiao Han
Summary: This paper introduces a weakly-supervised model using joint fully convolutional and graph convolutional networks for automated segmentation of pathology images. By utilizing image-level labels instead of pixel-wise annotations, the segmentation model's performance is improved. Experimental results demonstrate the effectiveness of this method in cancer region segmentation.
MEDICAL IMAGE ANALYSIS
(2021)
Article
Environmental Studies
Chong Niu, Kebo Ma, Xiaoyong Shen, Xiaoming Wang, Xiao Xie, Lin Tan, Yong Xue
Summary: Detecting areas where landslides or mudslides might occur is crucial, and previous studies have shown that convolutional neural networks (CNNs) outperform traditional methods for this task. However, CNNs typically focus on local features, which may be inefficient for recognizing complex landslide and mudslide scenes. To address this, this paper proposes an attention-enhanced region proposal network that integrates attentions into the architecture of state-of-the-art CNNs. Experimental results demonstrate that the proposed method outperforms unmodified CNNs in detecting non-landslides and non-mudslides, highlighting the importance of considering both local and global features in precise landslide and mudslide detection.
Article
Multidisciplinary Sciences
Sajib Saha, Janardhan Vignarajan, Shaun Frost
Summary: This paper presents a computationally efficient and memory efficient CNN-based system for automated detection of glaucoma. The system achieves high accuracy and speed while minimizing resource requirements. It performs well in classifying glaucomatous and non-glaucomatous images, making it suitable for integration into portable fundus cameras.
SCIENTIFIC REPORTS
(2023)
Article
Chemistry, Analytical
Yoshito Nagaoka, Tomo Miyazaki, Yoshihiro Sugaya, Shinichiro Omachi
Summary: This article introduces a novel text detection CNN architecture that is sensitive to text scale, proposing a method to solve the challenge of detecting small texts. Through the design of multi-resolution feature maps and receptive field size, the experimental results demonstrate the importance of receptive field size.
Article
Engineering, Electrical & Electronic
Fangyu Shen, Yanfei Wang, Chang Liu
Summary: This article introduces an unsupervised method for change detection, which improves change detection performance by enhancing the accuracy of pseudo-labels and designing a multi-scale feature fusion convolutional neural network. By eliminating the generation of differential images (DI) and directly obtaining results from original images, this method shows effectiveness in detecting changes in radar images with speckle noise.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2021)
Article
Computer Science, Artificial Intelligence
Ying-Chih Lo, I-Fang Chung, Shin-Ning Guo, Mei-Chin Wen, Chia-Feng Juang
Summary: This study uses deep learning techniques to automatically translate renal pathology images and detect glomeruli to improve the efficiency and accuracy of pathological diagnosis. Experimental results show that the automatic glomerulus detection method outperforms that manually labeled by physicians, effectively enhancing the efficiency of renal pathological diagnosis.
APPLIED SOFT COMPUTING
(2021)
Article
Computer Science, Artificial Intelligence
Rudiger Schmitz, Frederic Madesta, Maximilian Nielsen, Jenny Krause, Stefan Steurer, Rene Werner, Thomas Rosch
Summary: The study introduces multi-encoder fully-convolutional neural networks to integrate information from different spatial scales, enhancing the accuracy and efficiency of histopathologic diagnosis.
MEDICAL IMAGE ANALYSIS
(2021)
Article
Multidisciplinary Sciences
Di Lu, Shuli Cheng, Liejun Wang, Shiji Song
Summary: In this paper, a new deep learning-based method for change detection is proposed, which utilizes multi-scale feature fusion and distribution strategies to improve the accuracy of change region detection. Experimental results demonstrate that this method outperforms other comparative methods.
SCIENTIFIC REPORTS
(2022)
Article
Computer Science, Hardware & Architecture
Weidong Zhang, Xipeng Pan, Xiwang Xie, Lingqiao Li, Zimin Wang, Chu Han
Summary: This paper proposes an algorithm for color correction and adaptive contrast enhancement of underwater images, which compensates for lower color channels with dedicated fractions and applies an adaptive contrast enhancement algorithm and unsharp masking technique, achieving high-quality image enhancement.
COMPUTERS & ELECTRICAL ENGINEERING
(2021)
Article
Computer Science, Hardware & Architecture
Xiwang Xie, Weidong Zhang, Huadeng Wang, Lingqiao Li, Zhengyun Feng, Zhizhen Wang, Zimin Wang, Xipeng Pan
Summary: This study introduces a liver image segmentation method based on dynamic adaptive residual network, improving the accuracy and efficiency of liver image segmentation by optimizing the network structure and introducing methods such as conditional random fields.
COMPUTERS & ELECTRICAL ENGINEERING
(2021)
Article
Computer Science, Information Systems
Zhiwei Cao, Huihua Yang, Weijin Xu, Juan Zhao, Lingqiao Li, Xipeng Pan
Summary: The novel multiscale anchor-free (MSAF) region proposal network proposed in this paper is able to effectively obtain proposals for small-scale pedestrians and further improve detection performance with the use of a classifier.
WIRELESS COMMUNICATIONS & MOBILE COMPUTING
(2021)
Article
Chemistry, Analytical
Zhiwei Cao, Huihua Yang, Juan Zhao, Shuhong Guo, Lingqiao Li
Summary: This paper proposed a new method for multispectral pedestrian detection, which improved detection performance by fusing information from color and thermal streams and designed various fusion architectures to transfer feature information. Experimental results showed that the Halfway Fusion architecture performed the best among all architectures, and the new Multispectral Channel Feature Fusion (MCFF) module could adapt fused features in different modalities.
Article
Engineering, Electrical & Electronic
Longhao Zhang, Huihua Yang, Tian Qiu, Lingqiao Li
Summary: In this paper, a lightweight Generative Adversarial Networks framework named AP-GAN is proposed for efficient and high-fidelity video face swapping. The framework utilizes a U-Net based generator, a PE-aware discriminator, and a discriminator based perceptual loss leveraging multi-scale features to achieve precise control of facial attributes and consistency with the target face.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2022)
Article
Optics
Pengyou Fu, Yue Wen, Yuke Zhang, Lingqiao Li, Yanchun Feng, Lihui Yin, Huihua Yang
Summary: This study introduces a novel drug spectrum calibration model named SpectaTr based on the Transformer structure. The experimental results demonstrate that the proposed model can automatically extract features, is not dependent on pre-processing algorithms, and is insensitive to model hyperparameters. The SpectaTr model outperforms traditional methods and other deep learning models in terms of prediction accuracy.
JOURNAL OF INNOVATIVE OPTICAL HEALTH SCIENCES
(2022)
Article
Computer Science, Information Systems
Wenyi Zhao, Weidong Zhang, Xipeng Pan, Peixian Zhuang, Xiwang Xie, Lingqiao Li, Huihua Yang
Summary: Self-supervised visual representation learning aims to extract distinctive features from unlabeled datasets without labor-intensive data annotation. This study proposes an interleave sampling module inspired by LEGO bricks, which integrates the advantages of semantic and spatial relations to fully utilize unlabeled datasets. Experimental results demonstrate that this method outperforms state-of-the-art SSL methods on various datasets.
INFORMATION SCIENCES
(2022)
Article
Automation & Control Systems
Wei Li, Lingqiao Li, Huihua Yang
Summary: In this paper, a three-step Progressive Cross-domain Knowledge Distillation (PCdKD) paradigm is proposed for efficient unsupervised adaptive object detection on resource-constrained devices. The method progressively distills domain adaptive knowledge through image-to-image translation and a focal multi-domain discriminator, and utilizes adapted pseudo labels to retrain the teacher-student models. The proposed method boosts the transferability of the teacher model and enhances the student model for real-time applications.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2023)
Article
Engineering, Electrical & Electronic
Wenyi Zhao, Chongyi Li, Weidong Zhang, Lu Yang, Peixian Zhuang, Lingqiao Li, Kefeng Fan, Huihua Yang
Summary: In this paper, an efficient self-supervised framework called GLNet is proposed to address the issues of inadequate data utilization and feature-specificity in SSL. The framework incorporates novel sampling and ensemble learning strategies, and particularly embeds global contrastive and local location tasks to improve downstream task performance. Experimental results demonstrate that GLNet outperforms state-of-the-art SSL methods in terms of accuracy and training time.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2023)
Article
Computer Science, Artificial Intelligence
Lingqiao Li, Xiangkai Wang, Mengyu Yang, Hongwei Zhang
Summary: This study introduces feature fusion module and deformable convolution into the object detection network, improving the efficiency of shared bicycle detection. By constructing a shared bicycle dataset (SBD) for model training and testing, the improved method shows a 13% increase in mean average precision (mAP) compared to the original faster R-CNN, indicating its suitability for detecting shared bicycles. Experimental results on the Microsoft Common Objects (COCO) dataset also demonstrate a 5.8% higher mAP for this method compared to faster R-CNN before improvement.
IET IMAGE PROCESSING
(2023)
Article
Computer Science, Artificial Intelligence
Wenyi Zhao, Yibo Xu, Lingqiao Li, Huihua Yang
Summary: This study proposes a self-supervised method called SSL2 that is both efficient and effective. It increases the number of samples while maintaining connections between semantic features through a global and local sampling method, maintaining low computational complexity and mapping relationships between global comprehensive and local detailed features. An information retainer projection head (IRPH) is introduced to balance the information between detailed inconsistency and semantic consistency, and hybrid tasks are embedded into SSL2 to optimize the model. Extensive evaluations show that SSL2 outperforms existing self-supervised frameworks in computer vision tasks, achieving satisfactory performance on ImageNet linear classification and competitive results compared to other state-of-the-art methods.
KNOWLEDGE-BASED SYSTEMS
(2023)
Letter
Cell Biology
Chen Zhu, Yanfeng Shi, Jing Yu, Wenhao Zhao, Lingqiao Li, Jingxi Liang, Xiaolin Yang, Bing Zhang, Yao Zhao, Yan Gao, Xiaobo Chen, Xiuna Yang, Lu Zhang, Luke W. Guddat, Lei Liu, Haitao Yang, Zihe Rao, Jun Li
Article
Computer Science, Artificial Intelligence
Xipeng Pan, Jijun Cheng, Feihu Hou, Rushi Lan, Cheng Lu, Lingqiao Li, Zhengyun Feng, Huadeng Wang, Changhong Liang, Zhenbing Liu, Xin Chen, Chu Han, Zaiyi Liu
Summary: High throughput nuclear segmentation and classification is crucial in biological analysis, clinical diagnosis, and precision medicine. This study proposes the SMILE framework to address the challenges of nuclear heterogeneity, including imbalanced data distribution and diversified morphology characteristics. The proposed method achieves state-of-the-art performances on CoNSeP and MoNuSAC 2020 datasets by incorporating multi-task correlation attention and cost-sensitive learning strategies, as well as a novel post-processing step based on the marker-controlled watershed scheme.
MEDICAL IMAGE ANALYSIS
(2023)
Article
Spectroscopy
Fu Peng-you, Wen Yue, Zhang Yu-ke, Li Ling-qiao, Yang Hui-hua
Summary: The study presents a quantitative analysis modeling and model transfer framework based on convolutional neural networks and transfer learning to improve model prediction performance on one instrument and across instruments. Experimental results show that the proposed method outperforms PLS, SVM, and CNN in terms of model prediction performance.
SPECTROSCOPY AND SPECTRAL ANALYSIS
(2023)
Article
Spectroscopy
Lu Hao-xiang, Zhang Jing, Li Ling-qiao, Liu Zhen-bing, Yang Hui-hua, Feng Yan-chun, Yin Li-hui
Summary: Near-infrared spectroscopy is widely used in various fields such as drug detection, and its performance has been improved by combining machine learning and deep learning methods. The LAR-CARS method can effectively select characteristic wavelength points of the sample, leading to models with better robustness.
SPECTROSCOPY AND SPECTRAL ANALYSIS
(2021)