Search

Monocular depth estimation is essential in computer vision and robotics applications, including localization, mapping, and 3D object detection. In recent years, supervised learning algorithms that model large amounts of data have been successful in depth estimation. However, obtaining dense ground truth depth labels remains a challenge in supervised training. Therefore, unsupervised methods trained using monocular image sequences have gained wider attention. However, the depth estimation results of most existing models often produce blurred edges. Therefore, we propose various effective improvement strategies to construct a depth estimation network TL-Depth. (1) We propose a tower connection structure that utilizes convolutional processing to facilitate feature fusion, achieve precise semantic classification of pixels, and yield more accurate depth results. (2) We employ a Laplacian-filtering residual to focus on boundary information and enhance detailed results. (3) During the feature extraction stage, multiple pooling excitations are used by embedding them in the convolutional layer. This reduces redundant information while enhancing the network's feature extraction capability. The experimental results on the KITTI dataset and the Make3D dataset demonstrate that this method achieves good results compared to current methods. [ABSTRACT FROM AUTHOR]

The concept of episodic memory (EM) faces significant challenges by two claims: EM might not be a distinct memory system, and EM might be an epiphenomenon of a more general capacity for mental time travel (MTT). Nevertheless, the observations leading to these arguments do not preclude the existence of a mechanically and functionally distinct EM system. First, modular systems, like cognition, can have distinct subsystems that may not be distinguishable in the system's final output. EM could be such a subsystem, even though its effects may be difficult to distinguish from those of other subsystems. Second, EM could have a distinct and consistent low-level function, which is used in diverse high-level functions such as MTT. This article introduces the scenario construction framework, proposing that EM crucially rests on memory traces containing the gist of an episodic experience. During retrieval, EM traces trigger the reconstruction of semantic representations, which were active during the remembered episode, and are further enriched with semantic information, to generate a scenario of the past experience. This conceptualization of EM is consistent with studies on the neural basis of EM and resolves the two challenges while retaining the key properties associated with EM. This article is part of the theme issue 'Elements of episodic memory: lessons from 40 years of research'. [ABSTRACT FROM AUTHOR]

Knowledge graph-based recommender systems integrate user–item interactions with knowledge graph information through Graph Neural Networks (GNNs), demonstrating their effectiveness in addressing data sparsity and cold start issues. However, existing knowledge graph has more invalid connections; these noise connections are amplified during message aggregation in GNNs. To alleviate this problem, this paper proposes a contrastive learning filtering method based on multi-view generation. Traditional denoising methods mainly use edge perturbation or graph diffusion to randomly add or remove data, which are highly uncertain and often destroy the semantic information of the original data. We proposed the multi-view generation-based contrastive learning method can perform multiple different samples on the original data to create contrasting views. This ensures the semantic completeness of the original data while also providing rich semantic samples for contrastive learning. Specifically, the method first performs multiple rounds of data sampling from the user–item interaction graph and the knowledge graph to generate multiple views of items. Subsequently, knowledge embedding techniques are used to vectorize entities and relationships within these views. Finally, a contrastive task uses a designed loss to share semantic info among items, controlling node connections with item similarity. Through this method, the information of neighboring nodes (entities or relationships) can be propagated, the importance of neighboring nodes can be distinguished, and the recommendation result can be more precise. Extensive experiments on three benchmark datasets demonstrate that our proposed multi-view contrastive learning filtering approach significantly enhances performance in knowledge graph-based recommendation tasks. [ABSTRACT FROM AUTHOR]

Skeleton-based human interaction recognition is a formidable challenge that demands the capability to discern spatial, temporal, and interactive features. However, current research still faces some limitations in identifying spatial, temporal, and interaction features. Methods based on graph convolutional networks often prove to be insufficient in capturing interactive features and structural semantic information of skeletons. In order to solve this problem, we construct a Mutual-semantic Adjacency Matrix (MAM) by amalgamating the relative semantic attention of two skeleton sequences. This MAM was then integrated with the convolution of residual graphs to enhance the extraction of spatial and interaction features. We propose a novel interactive semantics neural network (ISNN) for skeleton-based human interaction recognition to hierarchically fuse MAM and structural semantic information. In addition, integrating the bone stream, we propose a two-stream Interactive Semantics Neural Network (2 s-ISNN). Experiments conducted with our models on two interaction datasets, NTU-RGB+D (mutual) and NTU-RGB+D 120 (mutual), demonstrate significantly improved recognition capabilities in comprehending human interactions. The source code is available at: https://github.com/czant1977/ISNN-master//. [ABSTRACT FROM AUTHOR]

Node classification is a crucial task for efficiently analyzing graph-structured data. Related semi-supervised methods have been extensively studied to address the scarcity of labeled data in emerging classes. However, two fundamental weaknesses hinder the performance: lacking the ability to mine latent semantic information between nodes, or ignoring to simultaneously capture local and global coupling dependencies between different nodes. To solve these limitations, we propose a novel semantic-enhanced graph neural networks with global context representation for semi-supervised node classification. Specifically, we first use graph convolution network to learn short-range local dependencies, which not only considers the spatial topological structure relationship between nodes, but also takes into account the semantic correlation between nodes to enhance the representation ability of nodes. Second, an improved Transformer model is introduced to reasoning the long-range global pairwise relationships, which has linear computational complexity and is particularly important for large datasets. Finally, the proposed model shows strong performance on various open datasets, demonstrating the superiority of our solutions. [ABSTRACT FROM AUTHOR]

The pervasive use of mobile big data has profoundly altered daily life, providing unprecedented convenience and efficiency. However, with the proliferation of mobile devices and the explosive growth of data volume, the issue of user privacy protection has become increasingly severe. Location information and semantic information, as the two core components of mobile big data, can directly reflect users' behavioral trajectories and thought dynamics, underscoring the importance of privacy protection. Although existing technologies can protect user data to a certain extent, traditional methods struggle to address increasingly sophisticated attack techniques in the face of evolving privacy threats. A comprehensive privacy protection scheme for mobile big data was proposed in this study, with a focus on two main areas: the privacy protection of location-based and semantic-based mobile big data. For location information protection, an uncertain graph model was employed to effectively resist combined attacks by jointly protecting the user layer and the location layer. For semantic information protection, a hypergraph clustering method was used to structurally protect the user layer and the semantic layer, enhancing the privacy security of semantic information. This study not only addresses existing gaps but also provides new solutions for mobile big data privacy protection, offering significant theoretical and practical value. [ABSTRACT FROM AUTHOR]

Ethereum smart contracts are a special type of computer programs. Once deployed on the blockchain, they cannot be modified. This presents a significant challenge to the security of smart contracts. Previous research has proposed static and dynamic detection tools to identify vulnerabilities in smart contracts. These tools check contract vulnerabilities based on predefined rules, and the accuracy of detection strongly depends on the design of the rules. However, the constant emergence of new vulnerability types and strategies for vulnerability protection leads to numerous false positives and false negatives by tools. To address this problem, we analyze the characteristics of vulnerabilities in smart contracts and the corresponding protection strategies. We convert the contracts' bytecode into an intermediate representation to extract semantic information of the contracts. Based on this semantic information, we establish a set of detection rules based on semantic facts and implement a vulnerability detection tool SafeCheck using static program analysis methods. The tool is used to detect six common types of vulnerabilities in smart contracts. We have extensively evaluated SafeCheck on real Ethereum smart contracts and compared it to other tools. The experimental results show that SafeCheck performs better in smart contract vulnerability detection compared to other typical tools, with a high F‐measure (up to 83.1%) for its entire dataset. [ABSTRACT FROM AUTHOR]

Aspect-level sentiment classification (ALSC) struggles with correctly trapping the aspects and corresponding sentiment polarity of a statement. Recently, several works have combined the syntactic structure and semantic information of sentences for more efficient analysis. The combination of sentence knowledge with graph neural networks has also proven effective at ALSC. However, there are still limitations on how to effectively fuse syntactic structure and semantic information when dealing with complex sentence structures and informal expressions. To deal with these problems, we propose an ALSC fusion network that combines graph neural networks with a simultaneous consideration of syntactic structure and semantic information. Specifically, our model is composed of a syntactic attention module and a semantic enhancement module. First, the syntactic attention module builds a dependency parse tree with the aspect term being the root, so that the model focuses better on the words closely related to the aspect terms, and captures the syntactic structure through a graph attention network. In addition, the semantic enhancement module generates the adjacency matrix through self-attention, which is processed by the graph convolutional network to obtain the semantic details. Lastly, the extracted features are merged to achieve sentiment classification. As verified by experiments, the model we propose can effectively enhance ALSC's behavior. [ABSTRACT FROM AUTHOR]

Abstract The task of image fusion for optical images and SAR images is to integrate valuable information from source images. Recently, owing to powerful generation, diffusion models, e.g., diffusion denoising probabilistic model and score-based diffusion model, are flourished in image processing, and there are some effective attempts in image fusion by scholars’ progressive explorations. However, the diffusion models for image fusion suffer from inevitable SAR speckle that seriously shelters from effective information in the same location of optical image. Besides, these methods integrate pixel-level features without information for high-level tasks, e.g., target detection and image classification, which leads fused images are insufficient and their application accuracies are low, for high-level tasks. To tackle these hurdles, we propose the semantic information guided diffusion posterior sampling for image fusion. Firstly, we employ the SAR-BM3D as preprocessing to despeckle. Then, the sampling model is established with fidelity, regularization and semantic information guidance term. The first two terms are obtained by the variational diffusion method via variational inference and first-order stochastic optimization. The last term is served by cross entropy loss between annotation and classification result from FLCNet we design. Finally, the experiments validate the feasibility and superiority of the proposed method on WHU-OPT-SAR dataset and DDHRNet dataset.

Abstract Printed Circuit Boards (PCBs) are key devices for the modern-day electronic technologies. During the production of these boards, defects may occur. Several methods have been proposed to detect PCB defects. However, detecting significantly smaller and visually unrecognizable defects has been a long-standing challenge. The existing two-stage and multi-stage object detectors that use only one layer of the backbone, such as Resnet’s third layer ( $$C_4$$ C 4 ) or fourth layer ( $$C_5$$ C 5 ), suffer from low accuracy, and those that use multi-layer feature maps extractors, such as Feature Pyramid Network (FPN), incur higher computational cost. Founded by these challenges, we propose a robust, less computationally intensive, and plug-and-play Attentive Context and Semantic Enhancement Module (ACASEM) for two-stage and multi-stage detectors to enhance PCB defects detection. This module consists of two main parts, namely adaptable feature fusion and attention sub-modules. The proposed model, ACASEM, takes in feature maps from different layers of the backbone and fuses them in a way that enriches the resulting feature maps with more context and semantic information. We test our module with state-of-the-art two-stage object detectors, Faster R-CNN and Double-Head R-CNN, and with multi-stage Cascade R-CNN detector on DeepPCB and Augmented PCB Defect datasets. Empirical results demonstrate improvement in the accuracy of defect detection.

Increasing research has focused on semantic communication, the goal of which is to convey accurately the meaning instead of transmitting symbols from the sender to the receiver. In this paper, we design a novel encoding and decoding semantic communication framework, which adopts the semantic information and the contextual correlations between items to optimize the performance of a communication system over various channels. On the sender side, the average semantic loss caused by the wrong detection is defined, and a semantic source encoding strategy is developed to minimize the average semantic loss. To further improve communication reliability, a decoding strategy that utilizes the semantic and the context information to recover messages is proposed in the receiver. Extensive simulation results validate the superior performance of our strategies over state-of-the-art semantic coding and decoding policies on different communication channels.

In recent years, the Internet of Things (IoT) has gradually developed applications such as collecting sensory data and building intelligent services, which has led to an explosion in mobile data traffic. Meanwhile, with the rapid development of artificial intelligence, semantic communication has attracted great attention as a new communication paradigm. However, for IoT devices, however, processing image information efficiently in real time is an essential task for the rapid transmission of semantic information. With the increase of model parameters in deep learning methods, the model inference time in sensor devices continues to increase. In contrast, the Pulse Coupled Neural Network (PCNN) has fewer parameters, making it more suitable for processing real-time scene tasks such as image segmentation, which lays the foundation for real-time, effective, and accurate image transmission. However, the parameters of PCNN are determined by trial and error, which limits its application. To overcome this limitation, an Improved Pulse Coupled Neural Networks (IPCNN) model is proposed in this work. The IPCNN constructs the connection between the static properties of the input image and the dynamic properties of the neurons, and all its parameters are set adaptively, which avoids the inconvenience of manual setting in traditional methods and improves the adaptability of parameters to different types of images. Experimental segmentation results demonstrate the validity and efficiency of the proposed self-adaptive parameter setting method of IPCNN on the gray images and natural images from the Matlab and Berkeley Segmentation Datasets. The IPCNN method achieves a better segmentation result without training, providing a new solution for the real-time transmission of image semantic information.

Printed Circuit Boards (PCBs) are key devices for the modern-day electronic technologies. During the production of these boards, defects may occur. Several methods have been proposed to detect PCB defects. However, detecting significantly smaller and visually unrecognizable defects has been a long-standing challenge. The existing two-stage and multi-stage object detectors that use only one layer of the backbone, such as Resnet's third layer ( C 4 ) or fourth layer ( C 5 ), suffer from low accuracy, and those that use multi-layer feature maps extractors, such as Feature Pyramid Network (FPN), incur higher computational cost. Founded by these challenges, we propose a robust, less computationally intensive, and plug-and-play Attentive Context and Semantic Enhancement Module (ACASEM) for two-stage and multi-stage detectors to enhance PCB defects detection. This module consists of two main parts, namely adaptable feature fusion and attention sub-modules. The proposed model, ACASEM, takes in feature maps from different layers of the backbone and fuses them in a way that enriches the resulting feature maps with more context and semantic information. We test our module with state-of-the-art two-stage object detectors, Faster R-CNN and Double-Head R-CNN, and with multi-stage Cascade R-CNN detector on DeepPCB and Augmented PCB Defect datasets. Empirical results demonstrate improvement in the accuracy of defect detection. [ABSTRACT FROM AUTHOR]

Knowledge graph embedding aims to map entities and relations into a low‐dimensional vector space for easy manipulation. However, frequent entities are updated more often than infrequent ones during training, leading to inadequate representation of the latter's embeddings, which, in turn, affects the model's overall performance in downstream tasks. To address this issue, we propose a semantic information guide and enhance (SGE) method. The SGE tackles the heterogeneity in the frequency of entities through semantic reconstruction and a guidance network. The semantic reconstruction strengthens the semantic relevance among all entities and connects entities with different frequencies in semantic space. The guidance network extends these connections to knowledge space, enhancing the expression abilities of infrequent entities' embeddings without compromising the embeddings of frequent entities. Experiments with four commonly used benchmark datasets show that the SGE method improves the performance of baseline models in most cases and that the method is model‐independent. [ABSTRACT FROM AUTHOR]

The significance of facial anti-spoofing algorithms in enhancing the security of facial recognition systems cannot be overstated. Current approaches aim to compensate for the model's shortcomings in capturing spatial information by leveraging spatio-temporal information from multiple frames. However, the additional branches to extract inter-frame details increases the model's parameter count and computational workload, leading to a decrease in inference efficiency. To address this, we have developed a robust and easily deployable facial anti-spoofing algorithm. In this paper, we propose Central Difference Convolution UNet++ (UCDCN), which takes advantage of central difference convolution and improves the characterization ability of invariant details in diverse environments. Particularly, we leverage domain knowledge from image segmentation and propose a multi-level feature fusion network structure to enhance the model's ability to capture semantic information which is beneficial for face anti-spoofing tasks. In this manner, UCDCN greatly reduces the number of model parameters as well as achieves satisfactory metrics on three popular benchmarks, i.e., Replay-Attack, Oulu-NPU and SiW. [ABSTRACT FROM AUTHOR]

The development of artificial intelligence and information communication technology has significantly propelled advancements in autonomous driving. The advent of autonomous driving has a profound impact on societal development and transportation methods. However, as intelligent systems, autonomous driving systems (ADSs) often make wrong judgements in specific scenarios, resulting in accidents. There is an urgent need for comprehensive testing and validation of ADSs. Metamorphic testing (MT) techniques have demonstrated effectiveness in testing ADSs. Nevertheless, existing testing methods primarily encompass relatively simple metamorphic relations (MRs) that only verify ADSs from a single perspective. To ensure the safety of ADSs, it is essential to consider the various elements of driving scenarios during the testing process. Therefore, this paper proposes MetaSem, a novel metamorphic testing method based on semantic information of autonomous driving scenes. Based on semantic information of the autonomous driving scenes and traffic regulations, we design 11 MRs targeting different scenario elements. Three transformation modules are developed to execute addition, deletion and replacement operations on various scene elements within the images. Finally, corresponding evaluation metrics are defined based on MRs. MetaSem automatically discovers inconsistent behaviours according to the evaluation metrics. Our empirical study on three advanced and popular autonomous driving models demonstrates that MetaSem not only efficiently generates visually natural and realistic scene images but also detects 11,787 inconsistent behaviours on three driving models. [ABSTRACT FROM AUTHOR]

The role and forms of information existence in physical and communication environments are analyzed. It is determined that information is an entity of the second order of complexity, assuming a non–atomic mode of existence, which takes shape in the processes of interaction and which can arise and disappear. By analogy with quantum mechanics, information can be represented as a superposition of possible semantic states of an information object. Information in the processes of storage and transmission exhibits the properties of a macro object, and, in the processes of information interaction with other information objects it exhibits wave properties. Information interaction is a stochastic process involving a physical source/receiver and transmission channels, the form of a message for its transmission and subsequent semantic interpretation in accordance with the domain, and means to determine the pragmatic value, as a result of which not only does the semantic state of the receiver change, but also an information object can also arise as a communicative form of presentation of information. Information interactions are essentially physical interactions that, in addition to the actual pair of interacting objects (including in the form of images), involve circumstances, such as laws and patterns, previous states, concomitant factors, established knowledge, and so on, which are "induced" by input information in the receiver environment. [ABSTRACT FROM AUTHOR]

Due to the vigorous development of big data, news topic text classification has received extensive attention, and the accuracy of news topic text classification and the semantic analysis of text are worth us to explore. The semantic information contained in news topic text has an important impact on the classification results. Traditional text classification methods tend to default the text structure to the sequential linear structure, then classify by giving weight to words or according to the frequency value of words, while ignoring the semantic information in the text, which eventually leads to poor classification results. In order to solve the above problems, this paper proposes a BiLSTM-GCN (Bidirectional Long Short-Term Memory and Graph Convolutional Network) hybrid neural network text classification model based on dependency parsing. Firstly, we use BiLSTM to complete the extraction of feature vectors in the text; Then, we employ dependency parsing to strengthen the influence of words with semantic relationship, and obtain the global information of the text through GCN; Finally, aim to prevent the overfitting problem of the hybrid neural network which may be caused by too many network layers, we add a global average pooling layer. Our experimental results show that this method has a good performance on the THUCNews and SogouCS datasets, and the F-score reaches 91.37% and 91.76% respectively. [ABSTRACT FROM AUTHOR]

As a driving force behind innovation, technological fusion has emerged as a prevailing trend in knowledge innovation. However, current research lacks the semantic analysis and identification of knowledge fusion across technological domains. To bridge this gap, we propose a strategy that combines the latent Dirichlet allocation (LDA) topic model and the Doc2vec neural network semantic model to identify fusion topics across various technology domains. Then, we fuse the semantic information of patents to measure the characteristics of fusion topics in terms of knowledge diversity, homogeneity and cohesion. Applying this method to a case study in the fields of digital medicine and machine learning, we identify six fusion topics from two technology domains, revealing two distinct trends: diffusion from the center to the periphery and clustering from the periphery to the center. The study shows that the fusion measure of topic-semantic granularity can reveal the variability of technology fusion processes at a profound level. The proposed research method will benefit scholars in conducting multi-domain technology fusion research and gaining a deeper understanding of the knowledge fusion process across technology domains from a semantic perspective. [ABSTRACT FROM AUTHOR]

With the advent and improvement of ontological dictionaries (WordNet, Babelnet), the use of synsets-based text representations is gaining popularity in classification tasks. More recently, ontological dictionaries were used for reducing dimensionality in this kind of representation (e.g., Semantic Dimensionality Reduction System (SDRS) (Vélez de Mendizabal et al., 2020)). These approaches are based on the combination of semantically related columns by taking advantage of semantic information extracted from ontological dictionaries. Their main advantage is that they not only eliminate features but can also combine them, minimizing (low-loss) or avoiding (lossless) the loss of information. The most recent (and accurate) techniques included in this group are based on using evolutionary algorithms to find how many features can be grouped to reduce false positive (FP) and false negative (FN) errors obtained. The main limitation of these evolutionary-based schemes is the computational requirements derived from the use of optimization algorithms. The contribution of this study is a new lossless feature reduction scheme exploiting information from ontological dictionaries, which achieves slightly better accuracy (specially in FP errors) than optimization-based approaches but using far fewer computational resources. Instead of using computationally expensive evolutionary algorithms, our proposal determines whether two columns (synsets) can be combined by observing whether the instances included in a dataset (e.g., training dataset) containing these synsets are mostly of the same class. The study includes experiments using three datasets and a detailed comparison with two previous optimization-based approaches. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Data Acquisition & Processing / Shu Ju Cai Ji Yu Chu Li is the property of Editorial Department of Journal of Nanjing University of Aeronautics & Astronautics and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Recently, there has been a significant increase in interest in cross-modal hashing technology. For hash code learning, most previous supervision methods use label information to create a similarity matrix in a straightforward manner. However, there are still the following challenges: (1) The asymmetric similarity matrix method only considers the similarity of labels, and the discriminant constraint in Hamming space is ignored; (2) there are optimization errors between the cross-modal semantic correlation of the Hamming space and the nonlinearity of the feature space; and (3) the cross-modal hash matrix is in a dynamic state during the optimization process, and the hash code is easily disturbed and there is bit uncertainty. To this end, we propose the Discriminative Latent Semantics-preserving Similarity Embedding Cross-modal Hashing (DLSSECH) method. Specifically, to reduce the quantization error, we introduce a non-asymmetric similarity decomposition based on orthogonal rotation bias. It can decompose the bitwise correlation of the learned hash code to capture more discriminative semantic information and compact hash code. In addition, to capture the cross-modal semantic correlation of nonlinear feature transformation and reduce quantization error, we propose the latent correlation error matrices based on orthogonal rotation decomposition. The model maintains the maximum difference in semantic dependencies between the projected data and its semantic representation. Finally, we use a sparse common hash matrix and non-asymmetric similarity decomposition factors to solve the uncertainty of the dynamic changes of the hash code. The effectiveness of DLSSECH has demonstrated through experiments on four cross-modal retrieval datasets, where it outperformed some state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Image classification is usually viewed as a visual recognition task and has extensive applications. Traditional efforts for image classification typical apply convolutional neural network (CNN) to identify the category of images. However, due to the limited receptive fields, it is difficult for CNN-based methods to model the global relations in images. This drawback leads to low classification accuracy and difficulty in handling complex and diverse image data. To model the global relationships, some researchers have applied Transformers to image classification tasks. However, to satisfy with the serialization and parallelization requirements of Transformers, the images need to be divided into equally sized and non-overlapping image patches, which breaks the local information between adjacent image blocks, which losses the spatiotemporal information of the whole image. Also, because of the limited prior knowledge of Transformers, models often need to be pre-trained on large-scale datasets, resulting in high computational complexity. To simultaneously model the spatiotemporal information between adjacent image blocks and fully utilize the global information of the images, this work propose a novel Spatiotemporal Convolutional Networks enhanced Transformer (SCN-Transformer) model for the basic image classification task. The SCN-Transformers approach can extract both local, global and spatiotemporal information between adjacent image blocks at a lower computational cost. The model comprises three components. In specific, the stacked Transformer modules capture the local correlations in the image, the SCN module fuses the local and spatiotemporal information between adjacent image blocks, and leverages long-range dependencies between different image blocks to enhance the representative capabilities of model's features, allowing the model to learn semantics from different dimensions. The classification module is responsible for the final image classification. The final experiments on the ImageNet 1K dataset demonstrate that the present model can outperform the existing mainstream image classification methods, and also achieve a competitive accuracy score of 83.7%, which confirms the competitiveness of the approach on large-scale image datasets. [ABSTRACT FROM AUTHOR]

Text classification is not only a prerequisite for natural language processing work, such as sentiment analysis and natural language reasoning, but is also of great significance for screening massive amounts of information in daily life. However, the performance of classification algorithms is always affected due to the diversity of language expressions, inaccurate semantic information, colloquial information, and many other problems. We identify three clues in this study, namely, core relevance information, semantic location associations, and the mining characteristics of deep and shallow networks for different information, to cope with these challenges. Two key insights about the text are revealed based on these three clues: key information relationship and word group inline relationship. We propose a novel attention feature fusion network, Attention Pyramid Transformer (APTrans), which is capable of learning the core semantic and location information from sentences using the above-mentioned two key insights. Specially, a hierarchical feature fusion module, Feature Fusion Connection (FFCon), is proposed to merge the semantic features of higher layers with positional features of lower layers. Thereafter, a Transformer-based XLNet network is used as the backbone to initially extract the long dependencies from statements. Comprehensive experiments show that APTrans can achieve leading results on the THUCNews Chinese dataset, AG News, and TREC-QA English dataset, outperforming most excellent pre-trained models. Furthermore, extended experiments are carried out on a self-built Chinese dataset theme analysis of teachers' classroom corpus. We also provide visualization work, further proving that APTrans has good potential in text classification work. [ABSTRACT FROM AUTHOR]

Visual Simultaneous Localization and Mapping (V-SLAM) is pivotal for precise positioning and mapping. However, visual data from crowd-sourced datasets often contains deficiencies that may lead to positioning errors. Despite existing optimization techniques, current algorithms do not adequately adapt to varied data in vehicle driving scenarios. To address this gap, this study introduces a novel SLAM framework (SLG-SLAM). This framework refines trajectories by integrating semantic information, laser point cloud, and global navigation satellite system (GNSS) data into V-SLAM. Initial trajectory estimates are made after filtering out dynamic targets and are subsequently refined with matched laser point clouds, then corrected for scale and direction using GNSS. The efficacy of this approach is assessed using four public datasets and one self-collected dataset, showing significant enhancements across all datasets. The proposed method reduces the mean absolute trajectory error by 43.50% on the KITTI dataset and 14.91% on the MVE dataset compared to the baseline. Unlike the baseline, which fails on three other datasets, the proposed method successfully performs localization and mapping. Additionally, compared to three other single-source methods (DynaSLAM, MCL, MVSLAM), the proposed method consistently outperforms, demonstrating its superior adaptability and effectiveness.

Abstract The significance of facial anti-spoofing algorithms in enhancing the security of facial recognition systems cannot be overstated. Current approaches aim to compensate for the model’s shortcomings in capturing spatial information by leveraging spatio-temporal information from multiple frames. However, the additional branches to extract inter-frame details increases the model’s parameter count and computational workload, leading to a decrease in inference efficiency. To address this, we have developed a robust and easily deployable facial anti-spoofing algorithm. In this paper, we propose Central Difference Convolution UNet++ (UCDCN), which takes advantage of central difference convolution and improves the characterization ability of invariant details in diverse environments. Particularly, we leverage domain knowledge from image segmentation and propose a multi-level feature fusion network structure to enhance the model’s ability to capture semantic information which is beneficial for face anti-spoofing tasks. In this manner, UCDCN greatly reduces the number of model parameters as well as achieves satisfactory metrics on three popular benchmarks, i.e., Replay-Attack, Oulu-NPU and SiW.

Introduction: Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) is a common sleep-related breathing disorder that significantly impacts the daily lives of patients. Currently, the diagnosis of OSAHS relies on various physiological signal monitoring devices, requiring a comprehensive Polysomnography (PSG). However, this invasive diagnostic method faces challenges such as data fluctuation and high costs. To address these challenges, we propose a novel data-driven Audio-SemanticMulti-Modalmodel forOSAHS severity classification (i.e., ASMM-OSA) based on patient snoring sound characteristics. Methods: In light of the correlation between the acoustic attributes of a patient's snoring patterns and their episodes of breathing disorders, we utilize the patient's sleep audio recordings as an initial screening modality. We analyze the audio features of snoring sounds during the night for subjects suspected of having OSAHS. Audio features were augmented via PubMedBERT to enrich their diversity and detail and subsequently classified for OSAHS severity using XGBoost based on the number of sleep apnea events. Results: Experimental results using the OSAHS dataset from a collaborative university hospital demonstrate that our ASMM-OSA audio-semanticmultimodal model achieves a diagnostic level in automatically identifying sleep apnea events and classifying the four-class severity (normal, mild, moderate, and severe) of OSAHS. Discussion: Our proposed model promises new perspectives for non-invasive OSAHS diagnosis, potentially reducing costs and enhancing patient quality of life. [ABSTRACT FROM AUTHOR]

This paper proposes incorporating linguistic semantic information into discourse relation recognition and constructing a Semantic Augmented Chinese Discourse Corpus (SACA) comprising 9546 adversative complex sentences. In adversative complex sentences, we suggest a quadruple (P, Q, R, Q β ) representing internal semantic elements, where the semantic opposition between Q and Q β forms the basis of the adversative relationship. P denotes the premise, and R represents the adversative reason. The overall annotation approach of this corpus follows the Penn Discourse Treebank (PDTB), except for the classification of senses. We combined insights from the Chinese Discourse Treebank (CDTB) and obtained eight sense categories for Chinese adversative complex sentences. Based on this corpus, we explore the relationship between sense classification and internal semantic elements within our newly proposed Chinese Adversative Discourse Relation Recognition (CADRR) task. Leveraging deep learning techniques, we constructed various classification models and the model that utilizes internal semantic element features, demonstrating their effectiveness and the applicability of our SACA corpus. Compared with pre-trained models, our model incorporates internal semantic element information to achieve state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Aspect-level sentiment classification is a fine-grained sentiment analysis that primarily focuses on predicting the sentiment polarity of aspects within a sentence. At present, many methods employ graph convolutional networks (GCN) to extract hidden semantic or syntactic information from sentences, achieving good results. However, these existing methods often overlook the relationships between multiple aspects within a sentence, treating aspects separately and thus neglecting the sentiment connections. To address this issue, this paper introduces a triple-channel graph attention network (TC-GAT) to capture semantics, syntax and multiple aspects dependencies information. In addition, a simple and effective fusion mechanism is proposed to comprehensively integrate these three types of information. Experiments are carried out on three commonly datasets, and the results verify the effectiveness of our proposed model. [ABSTRACT FROM AUTHOR]

Building extraction plays an important role in urban information analysis, which is helpful for urban planning and disaster monitoring. With the improvement of SAR resolution, rich detailed information in urban areas is revealed, but the discretized features also pose challenges for object detection. This paper addresses the problem of individual high-rise building extraction based on single high-resolution SAR image. Different from previous methods that require building facades to be presented in specific appearances, the proposed method is suitable for extraction of various types of high-rise buildings. After analyzing the SAR images of many types of high-rise buildings, we establish a unified high-rise building part model, on the basis of a scattering mechanism of building structures, to describe the facade characteristics of high-rise buildings, including a facade regularity part, facade bright line part, double bounce part, and their spatial topological relationships. A three-level high-rise building extraction framework, named HRBE-PM, is proposed based on the high-rise building part model. At the pixel level, a modified spot filter is used to extract bright spots and bright lines of different scales simultaneously to obtain salient features. At the structure level, building parts are generated based on the salient features according to semantic information. At the object level, spatial topological information between parts is introduced to generate building hypotheses. We define two indicators, completeness and compactness, to comprehensively evaluate each building hypothesis and select the optimal ones. After postprocessing, the final high-rise building extraction results are obtained. Experiments on two TerraSAR-X images show that the high-rise building extraction precision rate of the HRBE-PM method is above 85.29%, the recall rate is above 82.95%, and the F1-score is above 0.87. The results indicate that the HRBE-PM method can accurately extract individual high-rise buildings higher than 24 m in most dense scenes, regardless of building types. [ABSTRACT FROM AUTHOR]

In the field of Chinese-named entity recognition, recent research has sparked new interest by combining lexical features with character-based methods. Although this vocabulary enhancement method provides a new perspective, it faces two main challenges: firstly, using character-by-character matching can easily lead to conflicts during the vocabulary matching process. Although existing solutions attempt to alleviate this problem by obtaining semantic information about words, they still lack sufficient temporal sequential or global information acquisition; secondly, due to the limitations of dictionaries, there may be words in a sentence that do not match the dictionary. In this situation, existing vocabulary enhancement methods cannot effectively play a role. To address these issues, this paper proposes a method based on lexical information and spatial features. This method carefully considers the neighborhood and overlap relationships of characters in vocabulary and establishes global bidirectional semantic and temporal sequential information to effectively address the impact of conflicting vocabulary and character fusion on entity segmentation. Secondly, the attention score matrix extracted by the point-by-point convolutional network captures the local spatial relationship between characters without fused vocabulary information and characters with fused vocabulary information, aiming to compensate for information loss and strengthen spatial connections. The comparison results with the baseline model show that the SISF method proposed in this paper improves the F1 metric by 0.72%, 3.12%, 1.07%, and 0.37% on the Resume, Weibo, Ontonotes 4.0, and MSRA datasets, respectively. [ABSTRACT FROM AUTHOR]

The usual method of separating signal and noise principal components on the sole basis of their eigenvalues has evident drawbacks when semantically relevant information 'hides' in minor components, explaining a very small part of the total variance. This situation is common in biomedical experimentation when PCA is used for hypothesis generation: the multi-scale character of biological regulation typically generates a main mode explaining the major part of variance (size component), squashing potentially interesting (shape) components into the noise floor. These minor components should be erroneously discarded as noisy by the usual selection methods. Here, we propose a computational method, tailored for the chemical concept of 'titration', allowing for the unsupervised recognition of the potential signal character of minor components by the analysis of the presence of a negative linear relation between added noise and component invariance. [ABSTRACT FROM AUTHOR]

With the advent and improvement of ontological dictionaries (WordNet, Babelnet), the use of synsets-based text representations is gaining popularity in classification tasks. More recently, ontological dictionaries were used for reducing dimensionality in this kind of representation (e.g., Semantic Dimensionality Reduction System (SDRS) (Vélez de Mendizabal et al., 2020)). These approaches are based on the combination of semantically related columns by taking advantage of semantic information extracted from ontological dictionaries. Their main advantage is that they not only eliminate features but can also combine them, minimizing (low-loss) or avoiding (lossless) the loss of information. The most recent (and accurate) techniques included in this group are based on using evolutionary algorithms to find how many features can be grouped to reduce false positive (FP) and false negative (FN) errors obtained. The main limitation of these evolutionary-based schemes is the computational requirements derived from the use of optimization algorithms. The contribution of this study is a new lossless feature reduction scheme exploiting information from ontological dictionaries, which achieves slightly better accuracy (specially in FP errors) than optimization-based approaches but using far fewer computational resources. Instead of using computationally expensive evolutionary algorithms, our proposal determines whether two columns (synsets) can be combined by observing whether the instances included in a dataset (e.g., training dataset) containing these synsets are mostly of the same class. The study includes experiments using three datasets and a detailed comparison with two previous optimization-based approaches.

Chloride ion is severely harmful to reinforced concrete (RC) structures in marine environments. For maintaining the durability and safety of the designed RC structures, the determination of chloride ion concentration on concrete surfaces is critical. Currently, surface chloride ion concentration can be determined using empirical formulas and machine learning (ML) approaches. However, these approaches only rely on the numerical information within field records, disregarding valuable semantic and background information in the records, leading to low accuracy. Meanwhile, in splash environments, it presents a significant challenge to obtain chloride ion concentration records due to the complex environment and high costs involved. Therefore, based on limited field records of surface ion concentrations in splash environments, and utilizing a state-of-the-art language model (LM), this study proposes an LM-based information generation (LMIG) model to improve the accuracy of determination of surface chloride concentrations on RC structures. This paper utilizes the numerical and semantic information in 70 sets of field records to fine-tune the LMIG model and generates 200 sets of high-quality records. These records are then used to train ML algorithms for predicting chloride ion concentrations on concrete surfaces. After conducting comparative research, it was found that incorporating records generated by the LMIG model significantly enhances the accuracy of the ML algorithm. Specifically, the predictive accuracy using the random forest algorithm increased by 33.1%. Furthermore, this paper also conducts a comparative study on the LMIG model and the generative adversarial network (GAN)-assisted data-driven method. The results demonstrate that integrating semantic and numerical information into the LMIG model shows significant advantages in enhancing ML algorithms’ accuracy.

Image compression methods based on machine learning have achieved high rate-distortion performance. However, the reconstructions they produce suffer from blurring at extremely low bitrates (below 0.1 bpp), resulting in low perceptual quality. Although some methods attempt to reconstruct sharp images using Generative Adversarial Networks (GANs), reconstructing natural textures at low bitrates remains challenging. In this paper, we propose a novel image compression method that explicitly utilizes semantic information. Specifically, we send a semantic label map to the decoder, which takes it as input. This semantic information enables the decoder to reconstruct appropriate textures consistent with the corresponding semantic classes. Although semantic label maps can be compressed into relatively small data sizes using common methods (e.g., PNG), the data size is not negligible in an extremely low-rate setting. To address this problem, we propose simple yet effective label map compression strategies, including an autoregressive label map compressor. Our strategies significantly reduce the data size of the label map while maintaining the critical semantic information that allows the decoder to reconstruct realistic and suitable textures. By utilizing this data-efficient semantic information, our method can reconstruct realistic images even at an extremely low bitrate. As a result, the proposed method outperformed existing models, including a GAN-based model designed for low-rate settings and a state-of-the-art semantically guided method, in both quantitative evaluation and user studies. Furthermore, we analyzed the effect of semantic information by switching the input label map, confirming that the model synthesized textures appropriate to the given semantic labels.

Irony detection aims to analyze and identify language expressions containing irony in texts, which can assist text sentiment analysis and opinion mining. Hyperbolic expressions are the unique language characteristics in Chinese ironic texts that often highlight the critics’ ironic intentions. Existing methods of irony domain lexicon construction ignore the hyperbolic expressions in the construction process, resulting in incomplete semantic information contained in the lexicon. Considering this problem, this paper proposes a method for constructing a hyperbolic representation lexicon to mine the language feature and improve the accuracy of irony detection. Firstly, we select the candidate words and word pairs of hyperbolic representation by Word2Vec combined with the K-means++ algorithm. Then, the information entropy is calculated to measure the correlation between candidate words and texts, so the seed word set with a high correlation with texts is obtained. Finally, we expand the seed word set using WoBERT to capture the text’s deep semantic information. Thus, the generalization ability of the hyperbolic representation lexicon is improved. The experimental results show that combining the lexicon constructed in this paper can effectively mine the language features of ironic texts, thereby improving the accuracy of irony detection. The experimental indicators Acc and F1-score have an average improvement of about 2.38% and 2.29%, respectively.