检索结果-内蒙古大学图书馆

Low-Light Phase Retrieval With Implicit Generative Priors

IEEE TRANSACTIONS ON image processing 2024年 33卷 4728-4737页

作者： Manekar, Raunak Negrini, Elisa Pham, Minh Jacobs, Daniel Srivastava, Jaideep Osher, Stanley J. Miao, Jianwei Univ Minnesota Dept Comp Sci & Engn Minneapolis MN 55455 USA Univ Calif Los Angeles Dept Math Los Angeles CA 90095 USA Univ Calif Los Angeles Calif Nanosyst Inst Dept Phys & Astron Los Angeles CA 90095 USA

Phase retrieval (PR) is fundamentally important in scientific imaging and is crucial for nanoscale techniques like coherent diffractive imaging (CDI). Low radiation dose imaging is essential for applications involving radiation-sensitive samples. However, most PR methods struggle in low-dose scenarios due to high shot noise. Recent advancements in optical data acquisition setups, such as in-situ CDI, have shown promise for low-dose imaging, but they rely on a time series of measurements, making them unsuitable for single-image applications. Similarly, data-driven phase retrieval techniques are not easily adaptable to data-scarce situations. Zero-shot deep learning methods based on pre-trained and implicit generative priors have been effective in various imaging tasks but have shown limited success in PR. In this work, we propose low-dose deep image prior (LoDIP), which combines in-situ CDI with the power of implicit generative priors to address single-image low-dose phase retrieval. Quantitative evaluations demonstrate LoDIP's superior performance in this task and its applicability to real experimental scenarios.

关键词： Imaging Diffraction Noise Lighting image reconstruction Training Optimization deep generative models computational imaging phase retrieval inverse problems low light imaging low photon count zero-shot learning

来源：评论

学校读者我要写书评

暂无评论

deep learning Approach for a Machine-Human interface based on optical real-time Gesture Recognition for Automated Guided Vehicles

Deep Learning Approach for a Machine-Human interface based o...

引用

Conference on real-time processing of image, Depth, and Video Information

作者： Krishnakumar, Kiran Raj Gersmeier, Laura Harders, Leif Ole Hussmann, Stephan West Coast Univ Appl Sci Fritz Thiedemann Ring 20 D-25746 Heide Germany REINHOLZ Technol GmbH Kirchhoffstr 1A D-25524 Itzehoe Germany

ISBN: (纸本)9781510673199;9781510673182

The intersection of deep learning and programmable logic controllers (PLCs) can lead to innovative applications in automation. One of the exciting application areas are gesture-based control systems for Automated Guided Vehicles (AGVs). AGVs are used in various industries for material handling, logistics, warehouse automation, etc. Traditionally, these vehicles are controlled using predefined routes or remote controls, but with gesture-based control, operators can communicate more naturally and efficiently. The incorporation of YOLO-Pose in YOLO versions 7 and 8 has elevated the YOLO algorithm to a leading tool for creating gesture recognition models. The YOLO algorithm employs convolutional neural networks (CNN) to detect objects in real-time. These latest YOLO models offer significantly improved accuracy, speed, and reduced training times. This paper presents the comparative results of 2D gesture recognition transfer learning models created using the YOLO v5, v7, and v8 models, along with the steps taken to implement the model in a PLC-controlled AGV. Over 14,000 images were collected to build the models. A semi-automated approach was used to annotate them. Five models were created: two Keypoint models and three object detection models using transfer learning techniques with the same hyperparameters.

关键词： real-time image processing machine-human interface gesture recognition convolutional neural networks (CNN) automated guided vehicles machine learning deep-learning networks

来源：评论

学校读者我要写书评

暂无评论

deep learning-based time delay estimation for motion compensation in synthetic aperture sonars

引用

IET RADAR SONAR AND NAVIGATION 2024年第5期18卷 740-753页

作者： Chen, Shiping Chi, Cheng Zhang, Pengfei Wang, Peng Liu, Jiyuan Huang, Haining Chinese Acad Sci Inst Acoust Beijing Peoples R China Chinese Acad Sci Key Lab Sci & Technol Adv Underwater Acoust Signal Beijing Peoples R China Univ Chinese Acad Sci Beijing Peoples R China

Accurate and robust time delay estimation is crucial for synthetic aperture sonar (SAS) imaging. A two-step time delay estimation method based on displaced phase centre antenna (DPCA) micronavigation has been widely applied in motion estimation and compensation of SASs. However, the existing methods for time delay estimation are not sufficiently robust, which reduces the performance of SAS motion estimation. deep learning is currently one of the cutting-edge techniques and has brought about a remarkable progress in the field of underwater acoustic signal processing. In this study, a deep learning-based time delay estimation method is introduced to SAS motion estimation and compensation. The subband processing is first applied to obtain ambiguous time delays between adjacent pings from phases of SAS echoes. Then, a lightweight neural network is utilised to construct phase unwrapping. The model of the employed neural network is trained with simulation data and applied to real SAS data. The results of time delay estimation and motion compensation demonstrate that the proposed neural network-based method has much better performance than the two-step and joint-subband methods. In this study, a subband processing and lightweight neural network-based time delay estimation method is introduced to SAS motion estimation and compensation. The results of time delay estimation and motion compensation demonstrate that the proposed method has better performance than the reference method. In addition, this study provides an example of using simulation data to train a lightweight neural network and successfully applied to real ***

关键词： acoustic imaging acoustic signal processing motion compensation motion estimation synthetic aperture sonar

来源：评论

学校读者我要写书评

暂无评论

FLFusionSR: a fast and lightweight fusion and super-resolution network for infrared and visible images on edge devices

引用

JOURNAL OF real-time image processing 2024年第6期21卷 192页

作者： Xue, Weimin Liu, Yisha He, Guojian Wang, Fei Zhuang, Yan Dalian Maritime Univ Informat Sci & Technol Coll Dalian 116026 Peoples R China Dalian Maritime Univ Coll Artificial Intelligence Dalian 116026 Peoples R China Dalian Univ Technol Sch Control Sci & Engn Dalian 116024 Peoples R China

In real-world applications, the application of image fusion and super-resolution networks is limited by the computational burden and power consumption of edge devices. To resolve this problem, we propose a fast and lightweight network for infrared and visible image fusion and super-resolution called FLFusionSR. It directly learns end-to-end mapping from the original low-resolution infrared and visible images to a single high-resolution fused image. To meet the hardware requirements and real-time processing needs of edge devices, we adopt lightweight strategies in network design. To improve network performance without increasing computational complexity, we employ an auxiliary learning strategy during network training. The running analysis shows that our proposed method achieves 30FPS running speed on an edge device with 7.5 W power consumption, meeting the real-time processing requirements. Quantitative and visual comparison experiments on two datasets demonstrate that our proposed algorithm achieves state-of-the-art performance. The code is available at https://***/bearxwm/FLFusionSR.

关键词： Infrared and visible image fusion image fusion and super-resolution Auxiliary learning

来源：评论

学校读者我要写书评

暂无评论

SDHS-RLDNet: A real-time lightweight detection network for small-dense photovoltaic hot-spots

引用

JOURNAL OF real-time image processing 2025年第1期22卷 1-13页

作者： Hao, Shuai Li, Tong Ma, Xu Li, Tianqi Chang, Chi Qi, Tianrui Xian Univ Sci & Technol Xian 710054 Shaanxi Peoples R China

To ensure the safe and stable operation of photovoltaic power plants, it is crucial to conduct regular fault inspections on solar arrays. In a complex inspection environment, it is difficult to ensure the accuracy of detecting small and densely distributed photovoltaic hot-spot faults using traditional algorithms, and real-time detection is also challenging. To solve the above issues, a real-time lightweight detection network for small and dense photovoltaic hot-spots is proposed. Firstly, a multi-scale target extraction module is designed to enhance the feature extraction capability of the backbone network, which can effectively detect faults at different scales. Additionally, a small target prediction head is added to improve the detection performance of small hot-spot faults. Secondly, a dense object detection module is designed to enhance the positional information of hot-spot faults and effectively suppress background interference caused by complex backgrounds. Furthermore, to achieve network lightweighting, the method of knowledge distillation is adopted. By transferring the parameters of the teacher network to the student network, it simplifies the network parameters, improves model inference speed, and ensures real-time detection performance of the network. Finally, to verify the superiority of the proposed network, seven classical algorithms are selected for comparison experiments. The experimental results demonstrate that SDHS-RLDNet can accurately detect multi-scale hot-spot faults under various conditions with an accuracy rate reaching 86.6%.

关键词： deep learning real-time detection Lightweight network Fault inspection Hot-spot detection Knowledge distillation

来源：评论

学校读者我要写书评

暂无评论

Dilated-convolutional feature modulation network for efficient image super-resolution

引用

JOURNAL OF real-time image processing 2025年第2期22卷

作者： Wu, Lijun Li, Shan Chen, Zhicong Fuzhou Univ Coll Adv Mfg Quanzhou 362251 Peoples R China Fuzhou Univ Coll Phys & Informat Engn Fuzhou 350108 Peoples R China

In the field of image super-resolution (SR), deep learning-based models have achieved remarkable success. However, these models often face compatibility issues with low-power devices due to their computational and memory constraints. To address this challenge, numerous lightweight and efficient models have been proposed. While these models typically employ smaller convolutional kernels and shallower architectures to reduce parameter counts and computational complexity, they often neglect the importance of capturing global receptive fields. In this paper, we propose a simple yet effective deep network, termed the dilated-convolutional feature modulation network (DCFMN), to tackle these limitations. Specifically, we introduce a dilated separable modulation unit (DSMU) to aggregate spatial information from diverse large receptive fields. To complement the DSMU, which processes features from a long-range perspective, we further design a local feature enhancement module (LFEM) to extract local contextual information for effective channel fusion. Additionally, by leveraging reparameterization techniques, we ensure that the model incurs no additional computational overhead during inference. Extensive experimental results demonstrate that our DCFMN achieves competitive performance among existing efficient SR methods, while maintaining a compact model size and low computational complexity.

关键词： image super-resolution deep learning Lightweight Reparameterization

来源：评论

学校读者我要写书评

暂无评论

deep learning Algorithm Research and Performance Optimization of Financial Treasury Big Data Monitoring Platform

引用

INTERNATIONAL JOURNAL OF ADVANCED COMPUTER SCIENCE AND APPLICATIONS 2024年第6期15卷 322-331页

作者： Wang, Yanbing Ding, Ding Huishang Vocat Coll Dept Elect Informat Hefei 231201 Peoples R China AnHui Audit Coll Hefei 230601 Peoples R China

With the rapid development of information technology and the advent of the digital age, the management of fiscal treasury is facing unprecedented challenges and opportunities. In order to improve the efficiency and effectiveness of deep learning algorithms in the financial and treasury big data monitoring platform, this paper further studies the performance optimization methods of the model. This paper deeply studies deep learning algorithm research and performance optimization of financial Treasury big data monitoring platforms. This paper reviews the basic concepts, methods, and applications of deep learning and their application in the financial database big data monitoring platform. In the financial Treasury big data monitoring platform, deep learning algorithms are widely used in image recognition, natural language processing, recommendation systems and other fields. This article first conducts in-depth theoretical research on deep learning algorithms, including various neural network structures (such as convolutional neural network CNN, recurrent neural network RNN, etc.), optimization algorithms (such as gradient descent method and its variants), regularization techniques, etc. In addition, we also studied the practical applications of deep learning in fields such as image processing, natural language processing, and recommendation systems. In order to verify the effectiveness of deep learning algorithms in the financial and treasury big data monitoring platform, we designed corresponding experiments. These experiments include using deep learning algorithms for image recognition of financial documents, natural language processing, and building recommendation systems. We collected real fiscal treasury data as the experimental dataset and preprocessed and annotated the data.

关键词： deep learning financial database big data monitoring algorithm research performance optimization

来源：评论

学校读者我要写书评

暂无评论

Drone Attitude and Position Prediction via Stacked Hybrid deep learning Model for Massive MIMO Applications

引用

IEEE ACCESS 2024年 12卷 191377-191392页

作者： Al-Ahmadi, Abdullah Majmaah Univ Coll Engn Dept Elect Engn Majmaah 11952 Saudi Arabia

This paper presents a novel stacked hybrid deep learning model for real-time prediction of drone attitude and position, specifically designed to support applications in massive Multiple-Input Multiple-Output (MIMO) systems. The proposed model integrates Long Short-Term Memory (LSTM) networks, autoencoders, and attention mechanisms into a stacked architecture, enabling it to capture both short- and long-term dependencies as well as spatial features from complex drone flight data, including sensor readings and environmental conditions. This approach addresses limitations in traditional control methods, which struggle to maintain stability under noisy, dynamic environments. Extensive experiments were conducted to assess the model's accuracy and robustness across diverse flight scenarios. When compared to the Extended Kalman Filter (EKF), a traditional method for attitude estimation, the stacked hybrid model demonstrated superior performance, achieving significantly lower Mean Absolute Error (MAE) across all predicted parameters. Specifically, the stacked model achieved an average MAE of 0.0680, outperforming the EKF's MAE in predicting latitude, longitude, height, pitch, roll, and speed. The model's use of attention mechanisms further enhances its adaptability, allowing it to dynamically focus on the most relevant features within high-dimensional input data, improving prediction reliability in real-time scenarios.

关键词： Drones Predictive models Accuracy Sensors deep learning Estimation Long short term memory real-time systems image edge detection Gyroscopes Drone attitude prediction massive MIMO deep learning LSTM autoencoder attention mechanism

来源：评论

学校读者我要写书评

暂无评论

deep transfer learning-based computer vision for real-time harvest period classification and impurity detection of Porphyra haitnensis

引用

AQUACULTURE INTERNATIONAL 2024年第4期32卷 5171-5198页

作者： Gao, Zhenchang Huang, Jinxian Chen, Jiashun Shao, Tianya Ni, Hui Cai, Honghao Jimei Univ Sch Sci Dept Phys Xiamen Fujian Peoples R China Jimei Univ Coll Food & Biol Engn Xiamen Fujian Peoples R China Fujian Prov Key Lab Food Microbiol & Enzyme Engn Xiamen Fujian Peoples R China

Seaweed has attracted great attention as a healthy and nutritious food. Traditional seaweed processing plants mainly rely on manual visual inspection to identify and remove inferior seaweed. Accurate and rapid harvest period classification and impurity detection are key to improving productivity and processing speed in seaweed processing plants. Although many studies on seaweed have been conducted in laboratory environments, currently, the plants lack effective tools to obtain real-time and reliable information on seaweed quality. To address this challenge, the deep transfer learning-based computer vision was applied to identify inferior seaweeds, including those from the third harvest, fourth harvest, and impure seaweeds in this work. Specifically, YOLOv8 and YOLOv5 were utilized as base transfer learning models. By loading various pre-trained weight files, this study was able to automatically classify Porphyra haitnensis into four categories based on the harvest period and simultaneously detect four types of common impurities in it. Among the tested models, YOLOv8n-cls achieved the best trade-off in classifying the harvest period, with a Top-1 accuracy of 93.5%. This represented a significant improvement of 16% compared to the performance without transfer learning. The detection speed for a single image was 8.2 ms, and the model size was only 2.82 Mb. On the other hand, YOLOv8n exhibited outstanding performance in impurity detection, with a mean average precision of 99.14%, a single image detection speed of 4.3 ms, and a model size of 5.95 Mb. The results demonstrated the potential of YOLOv8 with transfer learning to objectively assist or even replace decision-making by assembly line workers. This study will not only enhance the quality control, production efficiency, and economic benefits of the seaweed processing industry but also drive the automation equipment and systems of seaweed-related enterprises towards greater intelligence and efficiency.

关键词： Transfer learning Computer vision YOLOv8 Seaweed Harvest period Impurity detection

来源：评论

学校读者我要写书评

暂无评论

Enhancing manufacturing process accuracy: A multidisciplinary approach integrating computer vision, machine learning, and control systems

引用

JOURNAL OF MANUFACTURING PROCESSES 2025年 142卷 453-467页

作者： Ramesh, Kaki Deshmukh, Sandip Ray, Tathagata Parimi, Chandu Birla Inst Sci & Technol Pilani Dept Mech Engn Hyderabad Campus Hyderabad 500078 Telangana India Birla Inst Sci & Technol Pilani Dept Comp Sci & Informat Syst Hyderabad Campus Hyderabad 500078 Telangana India Birla Inst Sci & Technol Pilani Dept Civil Engn Hyderabad Campus Hyderabad 500078 Telangana India

Manufacturing industries face significant challenges in producing high-quality, faultless products within limited timeframes. Conventional human-based inspection methods are still prone to errors and cannot guarantee precise component placement, potentially leading to product failures, user hazards, and substantial financial and reputational losses. This research presents a workflow to automate an inspection system that integrates computer vision, machine learning, image processing, and control systems to address these challenges. The proposed system employs a microcontroller and stepper motors to control a highly calibrated camera, enabling precise and efficient product inspection. At its core, the system utilizes the YOLOv5 model for object detection, specifically identifying hole marks and holes on products pre-assembly. This deep learning model was chosen for its real-time detection capabilities and high accuracy, achieving a mean Average Precision (mAP) of 0.95, which surpasses many current industry standards. Following object detection, advanced image processing techniques are applied to determine the precise position of detected features. Our approach achieves a notable error rate of 0.2 %, offering improvements over traditional inspection methods. Our system offers the potential to reduce inspection processing time and improve fault identification accuracy in real-time applications. Our research contributes to the field of industrial automation by introducing a seamless integration of state-of-the-art computer vision techniques with practical control systems. The system's modular design allows for easy adaptation to various manufacturing environments, benefiting industries with complex assembly processes, such as electronics, automotive manufacturing, etc. While the current implementation focuses on hole detection, future work will explore expanding the system's capabilities to identify a broader range of defects and adapt to different product types. This re

关键词： deep learning image processing Assembly Manufacturing industries Quality assurance

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：