检索结果-内蒙古大学图书馆

arXiv 2025年

作者： Mansourian, Amir M. Ahmadi, Rozhan Ghafouri, Masoud Babaei, Amir Mohammad Golezani, Elaheh Badali Ghamchi, Zeynab Yasamani Ramezanian, Vida Taherian, Alireza Dinashi, Kimia Miri, Amirali Kasaei, Shohreh Image Processing Lab Sharif University of Technology Iran

Deep Neural Networks (DNNs) have achieved notable performance in the fields of computer vision and natural language processing with various applications in both academia and industry. However, with recent advancements in DNNs and transformer models with a tremendous number of parameters, deploying these large models on edge devices causes serious issues such as high runtime and memory consumption. This is especially concerning with the recent large-scale foundation models, Vision-Language Models (VLMs), and Large Language Models (LLMs). Knowledge Distillation (KD) is one of the prominent techniques proposed to address the aforementioned problems using a teacher-student architecture. More specifically, a lightweight student model is trained using additional knowledge from a cumbersome teacher model. In this work, a comprehensive survey of knowledge distillation methods is proposed. This includes reviewing KD from different aspects: distillation sources, distillation schemes, distillation algorithms, distillation by modalities, applications of distillation, and comparison among existing methods. In contrast to most existing surveys, which are either outdated or simply update former surveys, this work proposes a comprehensive survey with a new point of view and representation structure that categorizes and investigates the most recent methods in knowledge distillation. This survey considers various critically important subcategories, including KD for diffusion models, 3D inputs, foundational models, transformers, and LLMs. Furthermore, existing challenges in KD and possible future research directions are discussed. Github page of the project: https://***/IPL-Sharif/KD_Survey © 2025, CC BY.

关键词： Deep neural networks

来源：评论

学校读者我要写书评

暂无评论

LEARNING CAUSAL RESPONSE REPRESENTATIONS THROUGH DIRECT EFFECT ANALYSIS

arXiv

引用

arXiv 2025年

作者： Durand, Homer Varando, Gherardo Camps-Valls, Gustau Image and Processing Lab Universitat de Valencia Valencia Spain

We propose a novel approach for learning causal response representations. Our method aims to extract directions in which a multidimensional outcome is most directly caused by a treatment variable. By bridging conditional independence testing with causal representation learning, we formulate an optimisation problem that maximises the evidence against conditional independence between the treatment and outcome, given a conditioning set. This formulation employs flexible regression models tailored to specific applications, creating a versatile framework. The problem is addressed through a generalised eigenvalue decomposition. We show that, under mild assumptions, the distribution of the largest eigenvalue can be bounded by a known F-distribution, enabling testable conditional independence. We also provide theoretical guarantees for the optimality of the learned representation in terms of signal-to-noise ratio and Fisher information maximisation. Finally, we demonstrate the empirical effectiveness of our approach in simulation and real-world experiments. Our results underscore the utility of this framework in uncovering direct causal effects within complex, multivariate settings. © 2025, CC BY-NC-SA.

关键词： Contrastive Learning

来源：评论

学校读者我要写书评

暂无评论

M-RAT: a Multi-grained Retrieval Augmentation Transformer for image Captioning 17th

M-RAT: a Multi-grained Retrieval Augmentation Transformer f...

引用

17th Asian Conference on Computer Vision, ACCV 2024

作者： Song, Jiayan Pan, Renjie Zhou, Jun Yang, Hua Institute of Image Communication and Network Engineering Shanghai Jiao Tong University Shanghai200240 China Shanghai Key Lab of Digital Media Processing and Transmission Shanghai200240 China

ISBN: (纸本)9789819609079

Current encoder-decoder methods for image captioning mai-nly consist of an object detection module (two-stage), or rely on big models with large-scale datasets to improve the effectiveness, which leads to increasing computation costs and cannot introduce new external knowledge. In this paper, we propose a novel end-to-end method Multi-grained Retrieval Augmentation Transformer (M-RAT) that innovatively fuses retrieved text derived from a changeable datastore with input visual feature through a Multi-modal Aligned Encoder, and introduce a specialized attention mechanism, Multi-MSA, to exploit both local and global interactions for delicate fine-grained details. Additionally, we enhance the decoder generation ability by employing low-level and high-level fused embeddings. Experiments demonstrate that M-RAT achieves comparable performance to state-of-the-art baselines with remarkable accuracy and details, as well as showing excellent domain adaptability for novel objects. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

关键词： Decoding

来源：评论

学校读者我要写书评

暂无评论

AMTrans:Auto-Correlation Multi-Head Attention Transformer for Infrared Spectral Deconvolution

引用

清华大学学报自然科学版（英文版） 2025年第3期30卷 1329-1341页

作者： Lei Gao Liyuan Cui Shuwen Chen Lizhen Deng Xiaokang Wang Xiaohong Yan Hu Zhu College of Electronic and Optical Engineering & College of Flexible Electronics(Future Technology) Nanjing University of Posts and TelecommunicationsNanjing 210003China Jiangsu Province Key Lab on Image Processing and Image Communication Nanjing University of Posts and TelecommunicationsNanjing 210003China National Engineering Research Center of Communication and Network Technology Nanjing University of Posts and TelecommunicationsNanjing 210003China School of Computer Science and Technology Hainan UniversityHaikou 570228China

Infrared spectroscopy analysis has found widespread applications in various fields due to advancements in technology and industry *** improve the quality and reliability of infrared spectroscopy signals,deconvolution is a crucial preprocessing *** by the transformer model,we propose an Auto-correlation Multi-head attention Transformer(AMTrans)for infrared spectrum sequence *** auto-correlation attention model improves the scaled dot-product attention in the *** utilizes attention mechanism for feature extraction and implements attention computation using the auto-correlation *** auto-correlation attention model is used to exploit the inherent sequence nature of spectral data and to effectively recovery spectra by capturing auto-correlation patterns in the *** proposed model is trained using supervised learning and demonstrates promising results in infrared spectroscopic *** comparing the experiments with other deconvolution techniques,the experimental results show that the method has excellent deconvolution performance and can effectively recover the texture details of the infrared spectrum.

关键词： spectroscopy spectral deconvolution transformer auto-correlation mechanism

来源：评论

学校读者我要写书评

暂无评论

A Turing Test for Artificial Nets devoted to model Human Vision

arXiv

引用

arXiv 2025年

作者： Vila-Tomás, Jorge Hernandez-Cámara, Pablo Li, Qiang Laparra, Valero Malo, Jesús Image Processing Lab Universitat de València València Paterna46980 Spain Image Processing Lab Universitat de València Spain TReNDS Georgia State Georgia Tech Emory United States

In this 2022 work2 3 we argued that, despite recent claims about successful modeling of the visual brain using artificial nets, the problem is far from being solved (even for low-level vision). First we introduce open issues such as where should we read from in ANNs in order to reproduce human behavior?, what should be the read-out mechanism?, this ad-hoc read-out is considered part of the brain model or not?, in order to understand the behavior of ANNs, should we use artificial psychophysics or artificial physiology?, in the case of ANNs, artificial experiments should literally match the experiments done with humans?. This means that there is a clear need of rigorous procedures for experimental tests for ANNs devoted to model the visual brain, and more generally, to understand ANNs devoted to generic vision tasks. Following our experience in using low-level facts from Quantitative Visual Neuroscience in image processing and computer vision, in this work we presented the idea of developing a low-level dataset compiling the basic spatio-temporal and chromatic facts that are known to happen in the retina-V1 pathway, and they are not currently available in existing databases such as BrainScore. In our illustrative results we checked the behavior of three recently proposed models with similar architecture: (1) A parametric model tuned via the psychophysical method of Maximum Differentiation [Malo & Simoncelli SPIE 15, Martinez et al. PLOS 18, Martinez et al. Front. Neurosci. 19], (2) A non-parametric model called PerceptNet tuned to maximize the correlation with human opinion about subjective distance between images [Hepburn et al. IEEE ICIP 19], and (3) A model with the same encoder as PerceptNet, but tuned for image segmentation (later published as Hernandez-Camara et al. ***. 23). Results on 10 compelling psycho/physio visual facts show that the first model (the parametric one) is the one with closer behavior to the human observers in terms of receptive

关键词： Brain

来源：评论

学校读者我要写书评

暂无评论

Analysis of multimodal fusion strategies in deep learning for ischemic stroke lesion segmentation on computed tomography perfusion data

引用

Multimedia Tools and Applications 2025年第10期84卷 7493-7518页

作者： Raju, Chintha Sri Pothu Neelapu, Bala Chakravarthy Laskar, Rabul Hussain Muhammad, Ghulam Speech and Image Processing Lab Department of ECE National Institute of Technology Silchar Assam Silchar788010 India Department of Biotechnology and Medical Engineering National Institute of Technology Rourkela Odisha Rourkela769008 India Department of Computer Engineering College of Computer and Information Sciences King Saud University Riyadh11543 Saudi Arabia

Stroke poses a significant risk to human life. Segmenting and immediately treating the stroke core stops its further development, therefore, enhancing the likelihood of survival. Convolutional neural networks (CNN) have been very successful in medical image segmentation, namely in the field of deep learning, and have produced the most advanced outcomes. Multi-modal images provide superior outcomes in the segmentation of stroke lesions compared to single-modal images. The integration of input from several modalities at various levels is crucial in determining performance and producing diverse outcomes in deep learning models that use multimodalities. Further investigation is required to explore the optimal methods for processing multimodal data in CNNs, the influence of fusion on CNN learning, and the effect of fusion strategies on lesions of varying sizes. To examine the impact of a multi-modal fusion method on lesion segmentation, we assessed four models using distinct fusion techniques, including early, late, bottleneck, and hierarchical fusions. This study discusses the various fusion procedures used in segmenting the lesion using computed tomography perfusion data. In addition, both quantitative and qualitative assessments, including deep feature analysis and feature similarity, were conducted to assess the impact of the fusion technique on the model’s performance. Furthermore, we examined the influence of fusion techniques on the size of the lesion. In addition, we analyzed the advantages and disadvantages of several multi-modal fusion systems. Our findings demonstrate that the bottleneck fusion technique got the highest dice score, 0.582, on the Ischemic Stroke Lesion Segmentation 2018 validation data as a result of its capacity to construct complex relationships across several modalities. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.

关键词： Convolutional neural networks

来源：评论

学校读者我要写书评

暂无评论

Fundamental Survey on Neuromorphic Based Audio Classification

arXiv

引用

arXiv 2025年

作者： Basu, Amlan Chaudhari, Pranav Di Caterina, Gaetano Neuromorphic Sensor Signal Processing Lab Centre for Image and Signal Processing Electrical and Electronic Engineering Department University of Strathclyde Glasgow United Kingdom

Audio classification is paramount in a variety of applications including surveillance, healthcare monitoring, and environmental analysis. Traditional methods frequently depend on intricate signal processing algorithms and manually crafted features, which may fall short in fully capturing the complexities of audio patterns. Neuromorphic computing, inspired by the architecture and functioning of the human brain, presents a promising alternative for audio classification tasks. This survey provides an exhaustive examination of the current state-of-the-art in neuromorphic-based audio classification. It delves into the crucial components of neuromorphic systems, such as Spiking Neural Networks (SNNs), memristors, and neuromorphic hardware platforms, highlighting their advantages in audio classification. Furthermore, the survey explores various methodologies and strategies employed in neuromorphic audio classification, including event-based processing, spike-based learning, and bio-inspired feature extraction. It examines how these approaches address the limitations of traditional audio classification methods, particularly in terms of energy efficiency, real-time processing, and robustness to environmental noise. Additionally, the paper conducts a comparative analysis of different neuromorphic audio classification models and benchmarks, evaluating their performance metrics, computational efficiency, and scalability. By providing a comprehensive guide for researchers, engineers and practitioners, this survey aims to stimulate further innovation and advancements in the evolving field of neuromorphic audio classification. © 2025, CC BY.

关键词： Benchmarking

来源：评论

学校读者我要写书评

暂无评论

IRSEnet: Differentially Private image Generation with Multi-Scale Feature Extraction and Residual Channel Attention 13

IRSEnet: Differentially Private Image Generation with Multi-...

引用

13th International Conference on Intelligent Control and Information processing, ICICIP 2025

作者： Li, Jiahao Wang, Zhongshuai Ghazali, Kamarul Hawari Bin Yan, Suqing Lan, Rushi Sun, Xiyan Luo, Xiaonan Guangxi Key Lab. of Image and Graphic Intelligent Processing Guilin University of Electronic Technology Guilin541004 China Centre for Advanced Industrial Technology University of Malaysia Pahang Al-Sultan Abdullah Pahang Pekan26600 Malaysia Int. Joint Research Lab. of Spatio-temporal Information and Intelligent Location Services Guilin University of Electronic Technology Guilin541004 China

ISBN: (纸本)9798331516147

Privacy-preserving image generation is particularly crucial in fields like healthcare, where data are both sensitive and limited. However, effective privacy preservation often compromises the visual quality and utility of the generated images due to privacy budget constraints. To address this issue, in this paper, We propose a novel network architecture, IRSEnet, which combines multi-scale feature extraction technology and residual channel attention mechanisms, aiming to enhance the visual quality of generated images and improve the performance of downstream classification tasks under differential privacy. The differential privacy mechanism ensures the security of sensitive data during training, while the multi-scale feature extraction module enhances feature extraction capabilities through parallel convolutional layers at multiple scales. Additionally, the channel attention module dynamically adjusts channel weights to focus on the most discriminative features. Experimental results demonstrate that this model significantly improves the utility of generated images and the accuracy of downstream classification tasks while preserving privacy. Future work will explore the application of this approach on larger datasets and across more diverse tasks. © 2025 IEEE.

关键词： Differential privacy

来源：评论

学校读者我要写书评

暂无评论

Fast and Accurate Video Analysis and Visualization of Classroom Activities Using Multiobjective Optimization of Extremely Low-parameter Models

引用

IEEE Access 2025年 13卷 81143-81166页

作者： Jatla, Venkatesh Teeparthi, Sravani Egala, Ugesh Celedon-Pattichis, Sylvia Pattichis, Marios S. University of New Mexico Image and Video Processing and Communications Lab Dept. of Electrical and Computer Engineering Albuquerque 87131 NM United States The University of Texas at Austin Department of Curriculum and Instruction Austin 78712-1293 TX United States

The paper considers the problem of video activity recognition in real-life collaborative classroom learning environments. Video analysis of real-life collaborative classroom learning environments faces significant challenges not encountered in current, advanced video recognition datasets. In collaborative learning environments, students are arranged in small groups where they interact within their group. Video analysis needs to deal with long-term activity recognition (of one hour or more session videos), detect multiple simultaneous activities, rapid transitions between activities, occlusions, and numerous individuals performing similar activities in the background that are not part of the group being analyzed. Developing ground truth datasets for analyzing complex video datasets is prohibitively expensive. We dramatically reduce the requirement for large ground truth datasets by creating separate, custom datasets for object detection and video activity *** then introduce a separable, extremely low-parameter system for video activity recognition that can be optimally trained using the derived datasets without the need for transfer learning from larger systems trained on large datasets. We further develop an interactive WebApp for visualizing the results over long video sessions. Overall, the extremely low-parameter activity classification model uses just 18.7K parameters for each activity, requiring 136.32 MB of memory. On a moderate GPU (RTX 5000), the activity classification model runs at an impressive 4,620 (154 x 30) frames per second. Our approach uses at least 1,000 fewer parameters than several well-established methods for video recognition. Our extremely low-parameter classifiers can process 90 minutes of video in just 26 seconds. Furthermore, our models are much easier to train, they are much faster, and outperform comparable methods. © 2013 IEEE.

关键词： educational video analysis extremely low-parameter neural networks fast inference Multi-objective optimization separable models video activity localization

来源：评论

学校读者我要写书评

暂无评论

Attention-based Fusion for Stroke Lesion Segmentation on Computed Tomography Perfusion Data

引用

ACM Transactions on Multimedia Computing, Communications and Applications 2025年第4期21卷

作者： Raju, Chintha Sri Pothu Laskar, Rabul Hussain Ali, Zulfiqar Muhammad, Ghulam Speech and Image Processing Lab Department of ECE National Institute of Technology Silchar Assam Silchar India University of Essex Colchester United Kingdom Department of Computer Engineering College of Computer and Information Sciences King Saud University Riyadh Riyadh Saudi Arabia

In recent times, stroke has emerged as a significant threat to humans, transforming affected brain tissue into core and penumbra regions. As the penumbra becomes irreversible over time, early core region segmentation is crucial. Automatic segmentation systems offer an efficient alternative to manual segmentation and aid radiologists in stroke lesion segmentation using computed tomography and Computed Tomography Perfusion (CTP) maps that comprise four parameter maps. This automatic segmentation is increasingly used in interactive, multimedia-based systems for diagnostic tools and AI-driven health applications. Top-performing models that follow the patch-processing approach suffer from high inference times. To incorporate effective feature extraction at image-level inferences, which reduces the inference time, we present a hybrid fusion technique that combines early and bottleneck fusion, leveraging two separate encoders for effective feature extraction. Moreover, fusing the information from various fusion methods arbitrarily may not yield optimal results. Consequently, we have introduced two attention modules, i.e., cross-modal attention and cross-fusion attention modules, designed for the effective integration of features derived from diverse modalities and multiple fusion strategies, respectively. The findings highlight a considerable reduction in computational time alongside achieving a comparable Dice score. Additionally, the incorporation of hybrid fusion and attention modules in the baseline notably increased the Dice score from 0.482 to 0.521 in the validation dataset and achieved 0.48 in the test dataset of ISLES 2018. It also demonstrates competitive performance compared to existing models while maintaining efficient prediction times. © 2025 Copyright held by the owner/author(s).

关键词： image segmentation

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：