Artificial Intelligence Review最新文献

Underwater object detection is a prerequisite for underwater robots to realize ocean exploration and autonomous grasping. However, underwater detection tasks face some inevitable interference factors, such as poor imaging quality, strong environment randomness, and high organism concealment. These phenomena will lead to strong underwater background interference and weak underwater object perception, which greatly aggravates the difficulty of underwater object detection. In order to deal with the above problems, we propose an unsupervised clustering optimization-based efficient attention (UCOEA). Different from the channel-wise strategy, cross-channel strategy and channel grouping strategy, we design a channel clustering strategy, which achieves autonomous dynamic screening of channel information by using the K-Means algorithm. Same types of channel information with high redundancy are learned uniformly to share the same operation. Different types of channel information with high specificity are learned independently to avoid channel noise information interference. Different from the single spatial strategy and multiple spatial strategy, we design a spatial clustering strategy, which achieves autonomous dynamic stripping of spatial information by using the EM algorithm. This strategy can extract multiple required spatial information at one time from different spatial locations. We further assign learnable weight parameters to distinguish dominant information and auxiliary information, which can alleviate spatial noise information interference. Our strategies can better balance additional cost overhead and information processing quality, which is crucial for the proposed attention to achieve fast and accurate underwater information calibration. In order to achieve high-precision and real-time underwater object detection, we propose a combined system of UCOEA underwater adapter and one-stage YOLO detector, which can efficiently detect small, medium and large targets at the same time. Extensive experiments demonstrate the effectiveness of our work. More importantly, we publish an underwater detection dataset DLMU2024 with low image continuity and high data diversity, which provides reliable support for the rapid development of underwater detection research. Our dataset is available at https://github.com/shenxin-dlmu/DLMU2024.

The rapid advancement of generative models has profoundly transformed the field of digital content creation, bringing unprecedented opportunities for media generation. However, the widespread adoption of this technology has also led to the emergence of highly realistic fake facial images and videos, which pose significant threats to public trust and societal security. To address the challenges of deepfake detection, this paper proposes a novel method based on Spatial-Frequency Feature Integration (SFFI), which effectively identifies fake content by combining spatial and frequency features of images. Additionally, to tackle the issue of class imbalance in the datasets, we propose an Authenticity-Aware Margin Loss (AAML). This loss function dynamically adjusts the decision boundary to enhance the model’s ability to recognize minority class samples. The proposed method was trained and evaluated on four challenging datasets: FaceForensics++, Celeb-DF v1, Celeb-DF v2, and the DeepFake Detection Challenge Preview, and compared against ten state-of-the-art methods. Experimental results demonstrate that the proposed method consistently outperforms all existing approaches across all datasets.

Time series forecasting is a critical task that provides key information for decision-making across various fields, such as economic planning, supply chain management, and medical diagnosis. After the use of traditional statistical methodologies and machine learning in the past, various fundamental deep learning architectures such as MLPs, CNNs, RNNs, and GNNs have been developed and applied to solve time series forecasting problems. However, the structural limitations caused by the inductive biases of each deep learning architecture constrained their performance. Transformer models, which excel at handling long-term dependencies, have become significant architectural components for time series forecasting. However, recent research has shown that alternatives such as simple linear layers can outperform Transformers. These findings have opened up new possibilities for using diverse architectures, ranging from fundamental deep learning models to emerging architectures and hybrid approaches. In this context of exploration into various models, the architectural modeling of time series forecasting has now entered a renaissance. This survey not only provides a historical context for time series forecasting but also offers comprehensive and timely analysis of the movement toward architectural diversification. By comparing and re-examining various deep learning models, we uncover new perspectives and present the latest trends in time series forecasting, including the emergence of hybrid models, diffusion models, Mamba models, and foundation models. By focusing on the inherent characteristics of time series data, we also address open challenges that have gained attention in time series forecasting, such as channel dependency, distribution shift, causality, and feature extraction. This survey explores vital elements that can enhance forecasting performance through diverse approaches. These contributions help lower entry barriers for newcomers by providing a systematic understanding of the diverse research areas in time series forecasting (TSF), while offering seasoned researchers broader perspectives and new opportunities through in-depth exploration of TSF challenges.

Specularity poses significant challenges in computer vision (CV), often leading to performance degradation in various tasks. Despite its importance, the CV field lacks a comprehensive review of specularity detection techniques. This survey addresses this gap by synthesizing diverse definitions of specularity and providing a unified framework to enhance consistency. It also presents a systematic review of traditional and deep learning-based methods for detecting specularity. Comparative experiments on a standardized dataset enable in-depth evaluation of each method, highlighting their strengths and limitations. The survey further provides structured insights and guidance for selecting appropriate methods across diverse scenarios. Through this, it identifies key areas for future research, aiming to support the development of more advanced detection models. By integrating diverse methodologies and quantitative analyzes, this survey contributes to a deeper understanding of current advancements and potential innovations in specularity detection.

Patents, which encapsulate crucial technical and legal information in text form and referenced drawings, present a rich domain for natural language processing (NLP). As NLP technologies evolve, large language models (LLMs) have demonstrated outstanding capabilities in general text processing and generation tasks. However, the application of LLMs in the patent domain remains under-explored and under-developed due to the complexity of patents, particularly their language and legal framework. Understanding the unique characteristics of patent documents and related research in the patent domain becomes essential for researchers to apply these tools effectively. Therefore, this paper aims to equip NLP researchers with the essential knowledge to navigate this complex domain efficiently. We introduce the relevant fundamental aspects of patents to provide solid background information. In addition, we systematically break down the structural and linguistic characteristics unique to patents and map out how NLP can be leveraged for patent analysis and generation. Moreover, we demonstrate the spectrum of text-based and multimodal patent-related tasks, including nine patent analysis and four patent generation tasks.

In this article, we presents a novel fuzzy neural network approach designed to address multi criteria decision making (MCDM) problems, specifically for selecting logistics service providers. The proposed decision making model integrates triangular fuzzy numbers (TFNs) with a triangular fuzzy Einstein weighted averaging (TFEWA) aggregation operator to enhance the decision making process under uncertainty. Initially, we discussed the concept of triangular fuzzy numbers, which allows for the representation of uncertain and imprecise data typically presented in real-world decision making environments. The operational laws, score function, and Hamming distance measures for TFNs are presented to ensure accurate handling of the fuzzy input data. The TFEWA aggregation operator, which is based on Einstein norms and plays a crucial role in aggregating expert opinions in the evaluation process. In the decision making process, we collect expert opinions regarding logistics service providers, expressed as TFNs, which are then processed through the fuzzy neural network model. After that, we apply the proposed decision making model to select the best logistics service providers. The TFEWA operator computes values at the hidden and output layers, and activation functions are applied to produce final output values. These outputs provide a ranked list of logistics service providers based on their overall performance across multiple criteria. The effectiveness of this novel approach is validated through a comparative analysis with existing MCDM methods. The results demonstrate that the triangular fuzzy neural network approach outperforms traditional methods in terms of flexibility, accuracy, and its ability to handle uncertain, fuzzy data. Our method provides a robust decision support system, capable of managing complex decision making tasks in logistics and other fields.

Sparse view 3D reconstruction remains challenging due to inherent data scale limitations. Mainstream sparse view 3D reconstruction algorithms based on the NeRF framework struggle to balance generation quality and real-time performance. Recently, the advent of 3D Gaussian Splatting technology has demonstrated remarkable results, becoming increasingly prominent in 3D scene representation and reconstruction. Exploring the application of 3D Gaussian Splatting technology for sparse view 3D reconstruction represents a promising research avenue. Based on this, our paper provides a comprehensive review of current sparse view 3D reconstruction methods leveraging 3D Gaussian Splatting, with an emphasis on extracting effective reconstruction information from input images and utilizing these data to generate realistic scenes efficiently and reliably. We then provide a detailed discussion on how the algorithm addresses issues such as artifacts and scale ambiguous, which are common challenges in this field. In the subsequent sections, we present both quantitative and qualitative comparisons of various sparse-view 3D reconstruction methods, roughly demonstrating the advantages of sparse view 3D Gaussian splatting methods in terms of reconstruction quality and efficiency. Furthermore, we analyze the potential applications of sparse view 3D Gaussian splatting methods. Finally, we identify the challenges faced by sparse-view 3D Gaussian splatting reconstruction and suggest potential solutions. We hope that our analysis will provide valuable insights for future research efforts.

Building management systems (BMSs) are increasingly integrating advanced machine learning (ML) and artificial intelligence (AI) capabilities to enhance operational efficiency and responsiveness. The transformation of BMSs involves a wide range of environmental, behavioural, economical and technical factors as well as optimum performance considerations in order to reach energy efficiency and for long term sustainability. Existing BMSs can only provide local adaptability by creating and managing information for a built asset lacking the capability to learn and adapt based on performance objectives. This research provides a comprehensive review of ML techniques in BMSs, with particular emphasis and demonstration of fast machine learning (FastML) techniques in a real-case study application. The study reviews optimization methods for ML algorithms, focusing on Long Short-Term Memory (LSTM) networks for energy consumption forecasting and exploring solutions that leverage hardware accelerators for low-latency and high-throughput processing. The High-Level Synthesis for Machine Learning (HLS4ML) framework facilitates deployment of fast machine learning models with BMSs, achieving substantial gains in hardware efficiency and inference speed in resource-constrained environments. Findings reveal that HLS4ML-optimized models maintain accuracy while offering computational efficiency through techniques like pruning and quantization, supporting real-time BMS applications. This research significantly contributes to the development of intelligent BMSs by integrating ML algorithms with advanced hardware solutions, ultimately improving energy management, occupant comfort, and safety in modern buildings.

Multi-modal representation learning is recognized for its comprehensive interpretation across diverse modalities. Although existing approaches have yielded favorable results, they face challenges in high-order information preservation and out-of-sample data generalization. To tackle these issues, we propose a scalable multi-modal representation learning networks framework, which aims to learn optimal modality-specific projection matrices to project multi-modal features to a shared representation space. Specifically, weight guided modality-wise and row-sparsity driven feature-wise measures are considered to achieve adaptively hierarchical feature selection from the original data. Then, within the unified latent representation space, we employ hypergraph embedding to preserve the intricate high-order local geometric structures within the modality-specific high-dimensional spaces. Finally, we propose a proximal operator-inspired network architecture to resolve the optimization objectives, streamlining the process of feature auto-weighted selection and representation learning. The experimental results highlight the effectiveness and superiority of the proposed method, while online testing on out-of-sample data further demonstrates robust generalization. The code of the proposed method is publicly available at: https://github.com/ZihanFang11/SMMRL.

Audio-visual zero-shot learning requires an understanding of the relationship between audio and visual information to determine unseen classes. Despite many efforts and significant progress in the field, many existing methods tend to focus on learning strong representations, neglecting the semantic consistency between audio and video as well as the inherent hierarchical structure of the data. To address these issues, we propose Learning Semantic Consistency for Audio-Visual Zero-shot Learning. Specifically, we employ an attention mechanism to enhance cross-modal information interactions, aiming to capture the semantic consistency between audio and visual data. Meanwhile, we introduce a hyperbolic space to model the hierarchical structure of the data itself. Moreover, the proposed approach includes a novel loss function that considers the relationships between input modalities, reducing the distance between features of different modalities. To evaluate the proposed method, we test it on three benchmark datasets (hbox {VGGSound-GZS}{{textrm{L}}^{cls}}), (hbox {UCF-GZS}{{textrm{L}}^{cls}}), and (hbox {ActivityNet-GZS}{{textrm{L}}^{cls}}). Extensive experimental results show that the proposed method achieves state-of-the-art performance on all three datasets. For example, on the (hbox {UCF-GZS}{{textrm{L}}^{cls}}) dataset, the harmonic mean is improved by 5.7%. Code and data available at https://github.com/ybyangjing/LSC-AVZSL.