检索结果-内蒙古大学图书馆

IEEE Transactions on artificial intelligence 2024年第10期5卷 5314-5324页

作者： Liu, Minjie Wang, Hongjian Yoon, Kuk-Jin Wang, Lin Artificial Intelligence Thrust Information Hub Guangzhou511466 China Tsinghua University Shenzhen International Graduate School Shenzhen518000 China Visual Intelligence Lab Department of Mechanical Engineering Daejeon366100 Korea Republic of Artificial Intelligence Thrust Guangzhou511466 China Department of Computer Science and Engineering SAR 999077 Hong Kong

Event cameras detect the intensity changes and produce asynchronous events with high dynamic range and no motion blur. Recently, several attempts have been made to superresolve the intensity images guided by events. However, these methods directly fuse the event and image features without distinguishing the modality difference and achieve image superresolution (SR) in multiple steps, leading to error-prone image SR results. Also, they lack quantitative evaluation of real-world data. In this article, we present an end-to-end framework, called event-guided image (EGI)-SR to narrow the modality gap and subtly integrate the event and RGB modality features for effective image SR. Specifically, EGI-SR employs three crossmodality encoders (CME) to learn modality-specific and modality-shared features from the stacked events and the intensity image, respectively. As such, EGI-SR can better mitigate the negative impact of modality varieties and reduce the difference in the feature space between the events and the intensity image. Subsequently, a transformer-based decoder is deployed to reconstruct the SR image. Moreover, we collect a real-world dataset, with temporally and spatially aligned events and color image pairs. We conduct extensive experiments on the synthetic and real-world datasets, showing EGI-SR favorably surpassing the existing methods by a large margin. © 2020 IEEE.

关键词： Image fusion

来源：评论

学校读者我要写书评

暂无评论

Small Object Few-Shot Segmentation for Vision-Based Industrial Inspection

引用

IEEE Transactions on Industrial Informatics 2025年第6期21卷 4388-4399页

作者： Zhang, Zilong Niu, Chang Zhao, Zhibin Zhang, Xingwu Chen, Xuefeng Xi'an Jiaotong University National Key Lab of Aerospace Power System and Plasma Technology Xi'an710049 China Foshan University School of Computer Science and Artificial Intelligence Foshan528000 China

Vision-based industrial inspection (VII) aims to locate defects quickly and accurately. Supervised learning under a close-set setting and industrial anomaly detection, as two common paradigms in VII, face different problems in practical applications. The former is that various and sufficient defects are difficult to obtain, while the latter is that specific defects cannot be located. To solve these problems, in this article, we focus on the few-shot semantic segmentation (FSS) method, which can locate unseen defects conditioned on a few annotations without retraining. Compared to common objects in natural images, the defects in VII are small. This brings two problems to current FSS methods: first, distortion of target semantics and second, many false positives for backgrounds. To alleviate these problems, we propose a small object few-shot segmentation (SOFS) model. The key idea for alleviating, first, is to avoid the resizing of the original image and correctly indicate the intensity of target semantics. SOFS achieves this idea via the nonresizing procedure and the prototype intensity downsampling of support annotations. To alleviate, second, we design an abnormal prior map in SOFS to guide the model in reducing false positives and propose a mixed normal dice loss to prevent the model from predicting false positives preferentially. SOFS can achieve FSS and few-shot anomaly detection determined by support masks. Diverse experiments substantiate the superior performance of SOFS. © 2025 IEEE. All rights reserved.

关键词： Semantic Segmentation

来源：评论

学校读者我要写书评

暂无评论

On the Emergence of Cross-Task Linearity in Pretraining-Finetuning Paradigm 41

On the Emergence of Cross-Task Linearity in Pretraining-Fine...

引用

41st International Conference on Machine Learning, ICML 2024

作者： Zhou, Zhanpeng Chen, Zijun Chen, Yilan Zhang, Bo Yan, Junchi School of Artificial Intelligence Department of Computer Science and Engineering MoE Lab of AI Shanghai Jiao Tong University Shanghai China Shanghai Artificial Intelligence Laboratory China Computer Science and Engineering University of California San Diego United States

The pretraining-finetuning paradigm has become the prevailing trend in modern deep *** this work, we discover an intriguing linear phenomenon in models that are initialized from a common pretrained checkpoint and finetuned on different tasks1, termed as Cross-Task Linearity (CTL).Specifically, we show that if we linearly interpolate the weights of two finetuned models, the features in the weight-interpolated model are often approximately equal to the linear interpolation of features in two finetuned models at each *** provide comprehensive empirical evidence supporting that CTL consistently occurs for finetuned models that start from the same pretrained *** conjecture that in the pretraining-finetuning paradigm, neural networks approximately function as linear maps, mapping from the parameter space to the feature *** on this viewpoint, our study unveils novel insights into explaining model merging/editing, particularly by translating operations from the parameter space to the feature ***, we delve deeper into the root cause for the emergence of CTL, highlighting the role of *** released our source code at https://***/zzp1012/Cross-Task-Linearity. Copyright 2024 by the author(s)

关键词： Deep learning

来源：评论

学校读者我要写书评

暂无评论

VSMCNN-dynamic summarization of videos using salient features from multi-CNN model

引用

Journal of Ambient intelligence and Humanized Computing 2023年第10期14卷 14071-14080页

作者： Nair, Madhu S. Mohan, Jesna Artificial Intelligence & Computer Vision Lab Department of Computer Science Cochin University of Science and Technology Kerala Kochi682022 India Department of Computer Science and Engineering Mar Baselios College of Engineering and Technology Nalanchira Kerala Thiruvananthapuram695015 India

A dynamic video summarization system detects key parts of the input video to generate its compact representation. The summaries can be used for efficient management of video data. This paper proposes an approach, Video summarization based on multi-CNN model (VSMCNN), that exploits major aspects of human cognition to generate meaningful summaries from videos. As the method focuses on dynamic summarization, the input video is divided into a set of shots. A multi-CNN model, which is a combination of different pre-trained models of CNN, is used for feature extraction from shots. The salient features are extracted from high dimensional feature vector using an unsupervised feature reduction technique applied in multiple subspaces to rank features in the vector. The distance measure between feature vectors is then thresholded to detect prime parts of the tested video. Experiments are performed on SumMe dataset and the results prove that our approach is successful in detecting portions of the tested video that has an essential message. The analysis shows that the method outperforms the state-of-the-art methods in the literature. Further evaluation on comparison with human-generated summaries in the ground truth proves the effectiveness of the proposed method. The paper also presents a detailed analysis to show which combination of pre-trained models of CNN is best suitable for generating dynamic summaries. © 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.

关键词： Video recording

来源：评论

学校读者我要写书评

暂无评论

Weakly Supervised Semantic Segmentation via Alternate Self-Dual Teaching

引用

IEEE Transactions on Image Processing 2023年 34卷 1-1页

作者： Zhang, Dingwen Li, Hao Zeng, Wenyuan Fang, Chaowei Cheng, Lechao Cheng, Ming-Ming Han, Junwei School of Automation Brain and Artificial Intelligence Laboratory Northwestern Polytechnical University Xi’an China School of Artificial Intelligence Xidian University Xi’an China Zhejiang Lab Hangzhou China College of Computer Science Nankai University TKLNDST Tianjin China

Weakly supervised semantic segmentation (WSSS) is a challenging yet important research field in vision community. In WSSS, the key problem is to generate high-quality pseudo segmentation masks (PSMs). Existing approaches mainly depend on the discriminative object part to generate PSMs, which would inevitably miss object parts or involve surrounding image background, as the learning process is unaware of the full object structure. In fact, both the discriminative object part and the full object structure are critical for deriving of high-quality PSMs. To fully explore these two information cues, we build a novel end-to-end learning framework, alternate self-dual teaching (ASDT), based on a dual-teacher single-student network architecture. The information interaction among different network branches is formulated in the form of knowledge distillation (KD). Unlike the conventional KD, the knowledge of the two teacher models would inevitably be noisy under weak supervision. Inspired by the Pulse Width (PW) modulation, we introduce a PW wave-like selection signal to alleviate the influence of the imperfect knowledge from either teacher model on the KD process. Comprehensive experiments on the PASCAL VOC 2012 and COCO-Stuff 10K demonstrate the effectiveness of the proposed ASDT framework, and new state-of-the-art results are achieved. IEEE

关键词： Distillation

来源：评论

学校读者我要写书评

暂无评论

DITune: A Reinforcement Learning Based Framework for Automated Database Index Selection 18

DITune: A Reinforcement Learning Based Framework for Automat...

引用

18th International Conference on Ubiquitous Information Management and Communication, IMCOM 2024

作者： Huang, Xiuqi Gao, Xiaofeng Chen, Guihai Shanghai Jiao Tong University MoE Key Lab of Artificial Intelligence Department of Computer Science and Engineering Shanghai China

ISBN: (纸本)9798350331011

Index selection is a crucial and essential aspect of database management systems (DBMS) aiming to optimize performance, and it is also one of the most time-consuming and laborious tasks for database administrators (DBAs). Therefore, automatic methods and tools are urgently needed to help DBAs choose a good index configuration for workloads. We propose a recommender system named DITune (Database Index Tuning) which can automatically analyze workloads, explore the potential benefits of indexes, and adjust the index configuration. We first implement a rule-based method to embed the database workload and data distribution feature. Then we utilize the Deep Deterministic Policy Gradient (DDPG) algorithm to generate index recommendations and construct index configuration. Moreover, we design a benchmark based on the standard TPC-C to validate the performance of our method under the varying workload patterns, query complexity, and running phase length. The experiment results show the effectiveness and robustness of DITune and indicate that the data-driven prediction outperforms those models based on cost estimation. © 2024 IEEE.

关键词： Reinforcement learning

来源：评论

学校读者我要写书评

暂无评论

On the Cut-Query Complexity of Approximating Max-Cut 51

On the Cut-Query Complexity of Approximating Max-Cut

引用

51st International Colloquium on Automata, Languages, and Programming, ICALP 2024

作者： Plevrakis, Orestis Ragavan, Seyoon Weinberg, S. Matthew Department of Computer Science Princeton University NJ United States Computer Science and Artificial Intelligence Lab Massachusetts Institute of Technology CambridgeMA United States

ISBN: (纸本)9783959773225

We consider the problem of query-efficient global max-cut on a weighted undirected graph in the value oracle model examined by [31]. Graph algorithms in this cut query model and other query models have recently been studied for various other problems such as min-cut, connectivity, bipartiteness, and triangle detection. Max-cut in the cut query model can also be viewed as a natural special case of submodular function maximization: on query S ⊆ V , the oracle returns the total weight of the cut between S and V \S. Our first main technical result is a lower bound stating that a deterministic algorithm achieving a c-approximation for any c > 1/2 requires Ω(n) queries. This uses an extension of the cut dimension to rule out approximation (prior work of [20] introducing the cut dimension only rules out exact solutions). Secondly, we provide a randomized algorithm with Õ(n) queries that finds a c-approximation for any c 1/2. Additionally, we show that any deterministic algorithm requires Ω(n2) queries to find an exact max-cut (enough to learn the entire graph). © Orestis Plevrakis, Seyoon Ragavan, and S. Matthew Weinberg.

关键词： Approximation algorithms

来源：评论

学校读者我要写书评

暂无评论

Light in the Dark: Cooperating Prior Knowledge and Prompt Learning for Low-Light Action Recognition 31

Light in the Dark: Cooperating Prior Knowledge and Prompt Le...

引用

31st IEEE International Conference on Image Processing Challenges and Workshops, ICIPCW 2024

作者： Liu, Qi Ye, Shu Wuhan University of Technology Hubei Key Lab of Transportation Internet of Things School of Computer Science and Artificial Intelligence China

ISBN: (纸本)9798331515942

Low-light environments present a significant challenge for action recognition. Due to insufficient lighting, video frames exhibit reduced detail and contrast, lacking the rich illumination information characteristic of standard RGB videos. Additionally, the dynamic range of cameras in low-light conditions is limited, making it difficult to capture changes in the video scenes. Thus, capturing motion information from low-light videos presents a significant challenge. We introduce a low-light action recognition model named Prior-knowledge Enhancement Prompt-learning Network (PEPN), which integrates video features based on illumination-invariant priors with large model motion description prompts. Unlike conventional methods, our method builds a feature bridge between low-light and standard videos through the use of illumination-invariant priors, addressing the issue of capturing lighting features in traditionally challenging low-light environments. Moreover, we enhance motion information using action descriptions. Specifically, we employ a pre-Trained large-scale language model as a knowledge engine to enhance lighting information with illumination-invariant priors and propose a multimodal training scheme. Furthermore, we align action descriptions with action features through joint understanding of video and text. We introduce an action-specific video enhancement module for priors with RGB features, which demonstrates superior performance on DARK-48 and INFAR datasets compared to previous models. © 2024 IEEE.

关键词： Motion capture

来源：评论

学校读者我要写书评

暂无评论

Disentangling Prosody Representations With Unsupervised Speech Reconstruction

引用

IEEE/ACM Transactions on Audio Speech and Language Processing 2024年 32卷 39-54页

作者： Qu, Leyuan Li, Taihao Weber, Cornelius Pekarek-Rosin, Theresa Ren, Fuji Wermter, Stefan Institute of Artificial Intelligence Zhejiang Lab Hangzhou311121 China Department of Informatics University of Hamburg Hamburg20148 Germany School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu611731 China

Human speech can be characterized by different components, including semantic content, speaker identity and prosodic information. Significant progress has been made in disentangling representations for semantic content and speaker identity in speech recognition and speaker verification tasks respectively. However, it is still an open challenging question to extract prosodic information because of the intrinsic association of different attributes, such as timbre and rhythm, and because of the need for supervised training schemes to achieve robust speech recognition. The aim of this article is to address the disentanglement of emotional prosody based on unsupervised reconstruction. Specifically, we identify, design, implement and integrate three crucial components in our proposed model Prosody2Vec: (1) a unit encoder that transforms speech signals into discrete units for semantic content, (2) a pretrained speaker verification model to generate speaker identity embeddings, and (3) a trainable prosody encoder to learn prosody representations. We first pretrain Prosody2Vec on unlabelled emotional speech corpora, then fine-tune the model on specific datasets to perform Speech Emotion Recognition (SER) and Emotional Voice Conversion (EVC) tasks. Both objective and subjective evaluations on the EVC task suggest that Prosody2Vec effectively captures general prosodic features that can be smoothly transferred to other emotional speech. In addition, our SER experiments on the IEMOCAP dataset reveal that the prosody features learned by Prosody2Vec are complementary and beneficial for the performance of widely used speech pretraining models and surpass the state-of-the-art methods when combining Prosody2Vec with HuBERT representations. Audio samples can be found on our demo website. © 2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

关键词： Speech recognition

来源：评论

学校读者我要写书评

暂无评论

The Detection of Abnormal Breathing Patterns in Cattle Using Optical Flow, Transfer Learning, and Ensemble Techniques 2

The Detection of Abnormal Breathing Patterns in Cattle Using...

引用

2nd International Conference of Intelligent Methods, Systems and Applications, IMSA 2024

作者： Elsayaad, Farida Daker, Mahmoud Atia, Ayman Helwan University HCI-LAB Faculty of Computers and Artificial Intelligence Giza Egypt Faculty of Computer Science Giza Egypt

ISBN: (纸本)9798350362633

In this paper, we focus on tracking cattle's respiratory behavior to detect abnormalities. We use a non-intrusive method which depends on recorded videos of cattle. These videos are filmed from the side and then used to produce breathing patterns. These breathing patterns are then classified as either normal or abnormal. Our method involves using YOLOv8 trained on the Animals with Attributes (AwA) dataset to automatically detect a point of interest on the cow's body. The movement of this point acts as a representative of the breathing behavior of the cow. This point is tracked using optical flow throughout the video, resulting in a breathing pattern. We tested the efficiency of 6 pretrained models in the task of classifying the resulting breathing patterns. We then combined the 3 highest-performing models which are VGG19, InceptionV3, and ResNet50 using ensemble techniques. After using ensemble techniques, an accuracy of 95.31%, a precision of 97.8%, and a recall of 92.27% was achieved, surpassing all individual models. © 2024 IEEE.

关键词： Invertebrates

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：