The remote sensing ships’fine-grained classification technology makes it possible to identify certain ship types in remote sensing images,and it has broad application prospects in civil and military fields.However,th...The remote sensing ships’fine-grained classification technology makes it possible to identify certain ship types in remote sensing images,and it has broad application prospects in civil and military fields.However,the current model does not examine the properties of ship targets in remote sensing images with mixed multi-granularity features and a complicated backdrop.There is still an opportunity for future enhancement of the classification impact.To solve the challenges brought by the above characteristics,this paper proposes a Metaformer and Residual fusion network based on Visual Attention Network(VAN-MR)for fine-grained classification tasks.For the complex background of remote sensing images,the VAN-MR model adopts the parallel structure of large kernel attention and spatial attention to enhance the model’s feature extraction ability of interest targets and improve the classification performance of remote sensing ship targets.For the problem of multi-grained feature mixing in remote sensing images,the VAN-MR model uses a Metaformer structure and a parallel network of residual modules to extract ship features.The parallel network has different depths,considering both high-level and lowlevel semantic information.The model achieves better classification performance in remote sensing ship images with multi-granularity mixing.Finally,the model achieves 88.73%and 94.56%accuracy on the public fine-grained ship collection-23(FGSC-23)and FGSCR-42 datasets,respectively,while the parameter size is only 53.47 M,the floating point operations is 9.9 G.The experimental results show that the classification effect of VAN-MR is superior to that of traditional CNNs model and visual model with Transformer structure under the same parameter quantity.展开更多
Fine-grained image classification, which aims to distinguish images with subtle distinctions, is a challenging task for two main reasons: lack of sufficient training data for every class and difficulty in learning dis...Fine-grained image classification, which aims to distinguish images with subtle distinctions, is a challenging task for two main reasons: lack of sufficient training data for every class and difficulty in learning discriminative features for representation. In this paper, to address the two issues, we propose a two-phase framework for recognizing images from unseen fine-grained classes, i.e., zeroshot fine-grained classification. In the first feature learning phase, we finetune deep convolutional neural networks using hierarchical semantic structure among fine-grained classes to extract discriminative deep visual features. Meanwhile, a domain adaptation structure is induced into deep convolutional neural networks to avoid domain shift from training data to test data. In the second label inference phase, a semantic directed graph is constructed over attributes of fine-grained classes. Based on this graph, we develop a label propagation algorithm to infer the labels of images in the unseen classes. Experimental results on two benchmark datasets demonstrate that our model outperforms the state-of-the-art zero-shot learning models. In addition, the features obtained by our feature learning model also yield significant gains when they are used by other zero-shot learning models, which shows the flexility of our model in zero-shot finegrained classification.展开更多
With the rapid development of the Internet of things and e-commerce, feature-based image retrieval and classification have become a serious challenge for shoppers searching websites for relevant product information. T...With the rapid development of the Internet of things and e-commerce, feature-based image retrieval and classification have become a serious challenge for shoppers searching websites for relevant product information. The last decade has witnessed great interest in research on content-based feature extraction techniques. Moreover, semantic attributes cannot fully express the rich image information. This paper designs and trains a deep convolutional neural network that the convolution kernel size and the order of network connection are based on the high efficiency of the filter capacity and coverage. To solve the problem of long training time and high resource share of deep convolutional neural network, this paper designed a shallow convolutional neural network to achieve the similar classification accuracy. The deep and shallow convolutional neural networks have data pre-processing, feature extraction and softmax classification. To evaluate the classification performance of the network, experiments were conducted using a public database Caltech256 and a homemade product image database containing 15 species of garment and 5 species of shoes on a total of 20,000 color images from shopping websites. Compared with the classification accuracy of combining content-based feature extraction techniques with traditional support vector machine techniques from 76.3% to 86.2%, the deep convolutional neural network obtains an impressive state-of-the-art classification accuracy of 92.1%, and the shallow convolutional neural network reached a classification accuracy of 90.6%. Moreover, the proposed convolutional neural networks can be integrated and implemented in other colour image database.展开更多
Accurate fine-grained geospatial scene classification using remote sensing imagery is essential for a wide range of applications.However,existing approaches often rely on manually zooming remote sensing images at diff...Accurate fine-grained geospatial scene classification using remote sensing imagery is essential for a wide range of applications.However,existing approaches often rely on manually zooming remote sensing images at different scales to create typical scene samples.This approach fails to adequately support the fixed-resolution image interpretation requirements in real-world scenarios.To address this limitation,we introduce the million-scale fine-grained geospatial scene classification dataset(MEET),which contains over 1.03 million zoom-free remote sensing scene samples,manually annotated into 80 fine-grained categories.In MEET,each scene sample follows a scene-in-scene layout,where the central scene serves as the reference,and auxiliary scenes provide crucial spatial context for fine-grained classification.Moreover,to tackle the emerging challenge of scene-in-scene classification,we present the context-aware transformer(CAT),a model specifically designed for this task,which adaptively fuses spatial context to accurately classify the scene samples.CAT adaptively fuses spatial context to accurately classify the scene samples by learning attentional features that capture the relationships between the center and auxiliary scenes.Based on MEET,we establish a comprehensive benchmark for fine-grained geospatial scene classification,evaluating CAT against 11 competitive baselines.The results demonstrate that CAT significantly outperforms these baselines,achieving a 1.88%higher balanced accuracy(BA)with the Swin-Large backbone,and a notable 7.87%improvement with the Swin-Huge backbone.Further experiments validate the effectiveness of each module in CAT and show the practical applicability of CAT in the urban functional zone mapping.The source code and dataset will be publicly available at https://jerrywyn.github.io/project/MEET.html.展开更多
In this paper,we propose hierarchical attention dual network(DNet)for fine-grained image classification.The DNet can randomly select pairs of inputs from the dataset and compare the differences between them through hi...In this paper,we propose hierarchical attention dual network(DNet)for fine-grained image classification.The DNet can randomly select pairs of inputs from the dataset and compare the differences between them through hierarchical attention feature learning,which are used simultaneously to remove noise and retain salient features.In the loss function,it considers the losses of difference in paired images according to the intra-variance and inter-variance.In addition,we also collect the disaster scene dataset from remote sensing images and apply the proposed method to disaster scene classification,which contains complex scenes and multiple types of disasters.Compared to other methods,experimental results show that the DNet with hierarchical attention is robust to different datasets and performs better.展开更多
Bird monitoring and protection are essential for maintaining biodiversity,and fine-grained bird classification has become a key focus in this field.Audio-visual modalities provide critical cues for this task,but robus...Bird monitoring and protection are essential for maintaining biodiversity,and fine-grained bird classification has become a key focus in this field.Audio-visual modalities provide critical cues for this task,but robust feature extraction and efficient fusion remain major challenges.We introduce a multi-stage fine-grained audiovisual fusion network(MSFG-AVFNet) for fine-grained bird species classification,which addresses these challenges through two key components:(1) the audiovisual feature extraction module,which adopts a multi-stage finetuning strategy to provide high-quality unimodal features,laying a solid foundation for modality fusion;(2) the audiovisual feature fusion module,which combines a max pooling aggregation strategy with a novel audiovisual loss function to achieve effective and robust feature fusion.Experiments were conducted on the self-built AVB81and the publicly available SSW60 datasets,which contain data from 81 and 60 bird species,respectively.Comprehensive experiments demonstrate that our approach achieves notable performance gains,outperforming existing state-of-the-art methods.These results highlight its effectiveness in leveraging audiovisual modalities for fine-grained bird classification and its potential to support ecological monitoring and biodiversity research.展开更多
Intelligent vehicle applications provide convenience but raise privacy and security concerns.Misuse of sensitive data,including vehicle location,and facial recognition information,poses a threat to user privacy.Hence,...Intelligent vehicle applications provide convenience but raise privacy and security concerns.Misuse of sensitive data,including vehicle location,and facial recognition information,poses a threat to user privacy.Hence,traffic classification is vital for promptly overseeing and controlling applications with sensitive information.In this paper,we propose ETNet,a framework that combines multiple features and leverages self-attention mechanisms to learn deep relationships between packets.ET-Net employs a multisimilarity triplet network to extract features from raw bytes,and exploits self-attention to capture long-range dependencies within packets in a session and contextual information features.Additionally,we utilizing the loss function to more effectively integrate information acquired from both byte sequences and their corresponding lengths.Through simulated evaluations on datasets with similar attributes,ET-Net demonstrates the ability to finely distinguish between nine categories of applications,achieving superior results compared to existing methods.展开更多
In this paper, we introduce an image dataset for fine-grained classification of dog breeds: the Tsinghua Dogs Dataset. It is currently the largest dataset for fine-grained classification of dogs, including 130 dog bre...In this paper, we introduce an image dataset for fine-grained classification of dog breeds: the Tsinghua Dogs Dataset. It is currently the largest dataset for fine-grained classification of dogs, including 130 dog breeds and 70,428 real-world images. It has only one dog in each image and provides annotated bounding boxes for the whole body and head. In comparison to previous similar datasets, it contains more breeds and more carefully chosen images for each breed. The diversity within each breed is greater,with between 200 and 7000+ images for each breed.Annotation of the whole body and head makes the dataset not only suitable for the improvement of finegrained image classification models based on overall features, but also for those locating local informative parts. We show that dataset provides a tough challenge by benchmarking several state-of-the-art deep neural models. The dataset is available for academic purposes at https://cg.cs.tsinghua.edu.cn/ThuDogs/.展开更多
The value of grape cultivars varies.The use of a mixture of cultivars can negate the benefits of improved cultivars and hamper the protection of genetic resources and the identification of new hybrid cultivars.Classif...The value of grape cultivars varies.The use of a mixture of cultivars can negate the benefits of improved cultivars and hamper the protection of genetic resources and the identification of new hybrid cultivars.Classifying cultivars based on their leaves is therefore highly practical.Transplanted grape seedlings take years to bear fruit,but leaves mature in months.Foliar morphology differs among cultivars,so identifying cultivars based on leaves is feasible.Different cultivars,however,can be bred from the same parents,so the leaves of some cultivars can have similar morphologies.In this work,a pyramid residual convolution neural network was developed to classify images of eleven grape cultivars.The model extracts multi-scale feature maps of the leaf images through the convolution layer and enters them into three residual convolution neural networks.Features are fused by adding the value of the convolution kernel feature matrix to enhance the attention on the edge and center regions of the leaves and classify the images.The results indicated that the average accuracy of the model was 92.26%for the proposed leaf dataset.The proposed model is superior to previous models and provides a reliable method for the fine-grained classification and identification of plant cultivars.展开更多
In recent decades,the proliferation of email communication has markedly escalated,resulting in a concomitant surge in spam emails that congest networks and presenting security risks.This study introduces an innovative...In recent decades,the proliferation of email communication has markedly escalated,resulting in a concomitant surge in spam emails that congest networks and presenting security risks.This study introduces an innovative spam detection method utilizing the Horse Herd Optimization Algorithm(HHOA),designed for binary classification within multi⁃objective framework.The method proficiently identifies essential features,minimizing redundancy and improving classification precision.The suggested HHOA attained an impressive accuracy of 97.21%on the Kaggle email dataset,with precision of 94.30%,recall of 90.50%,and F1⁃score of 92.80%.Compared to conventional techniques,such as Support Vector Machine(93.89%accuracy),Random Forest(96.14%accuracy),and K⁃Nearest Neighbours(92.08%accuracy),HHOA exhibited enhanced performance with reduced computing complexity.The suggested method demonstrated enhanced feature selection efficiency,decreasing the number of selected features while maintaining high classification accuracy.The results underscore the efficacy of HHOA in spam identification and indicate its potential for further applications in practical email filtering systems.展开更多
Visual diagnosis of skin cancer is challenging due to subtle inter-class similarities,variations in skin texture,the presence of hair,and inconsistent illumination.Deep learning models have shown promise in assisting ...Visual diagnosis of skin cancer is challenging due to subtle inter-class similarities,variations in skin texture,the presence of hair,and inconsistent illumination.Deep learning models have shown promise in assisting early detection,yet their performance is often limited by the severe class imbalance present in dermoscopic datasets.This paper proposes CANNSkin,a skin cancer classification framework that integrates a convolutional autoencoder with latent-space oversampling to address this imbalance.The autoencoder is trained to reconstruct lesion images,and its latent embeddings are used as features for classification.To enhance minority-class representation,the Synthetic Minority Oversampling Technique(SMOTE)is applied directly to the latent vectors before classifier training.The encoder and classifier are first trained independently and later fine-tuned end-to-end.On the HAM10000 dataset,CANNSkin achieves an accuracy of 93.01%,a macro-F1 of 88.54%,and an ROC–AUC of 98.44%,demonstrating strong robustness across ten test subsets.Evaluation on the more complex ISIC 2019 dataset further confirms the model’s effectiveness,where CANNSkin achieves 94.27%accuracy,93.95%precision,94.09%recall,and 99.02%F1-score,supported by high reconstruction fidelity(PSNR 35.03 dB,SSIM 0.86).These results demonstrate the effectiveness of our proposed latent-space balancing and fine-tuned representation learning as a new benchmark method for robust and accurate skin cancer classification across heterogeneous datasets.展开更多
Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in comp...Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in complex backgrounds,small target objects,and limited training data,leading to poor recognition.Fine-grained images exhibit“small inter-class differences,”and while second-order feature encoding enhances discrimination,it often requires dual Convolutional Neural Networks(CNN),increasing training time and complexity.This study proposes a model integrating discriminative region localization and efficient second-order feature encoding.By ranking feature map channels via a fully connected layer,it selects high-importance channels to generate an enhanced map,accurately locating discriminative regions.Cropping and erasing augmentations further refine recognition.To improve efficiency,a novel second-order feature encoding module generates an attention map from the fourth convolutional group of Residual Network 50 layers(ResNet-50)and multiplies it with features from the fifth group,producing second-order features while reducing dimensionality and training time.Experiments on Caltech-University of California,San Diego Birds-200-2011(CUB-200-2011),Stanford Car,and Fine-Grained Visual Classification of Aircraft(FGVC Aircraft)datasets show state-of-the-art accuracy of 88.9%,94.7%,and 93.3%,respectively.展开更多
Arrhythmias are a frequently occurring phenomenon in clinical practice,but how to accurately dis-tinguish subtle rhythm abnormalities remains an ongoing difficulty faced by the entire research community when conductin...Arrhythmias are a frequently occurring phenomenon in clinical practice,but how to accurately dis-tinguish subtle rhythm abnormalities remains an ongoing difficulty faced by the entire research community when conducting ECG-based studies.From a review of existing studies,two main factors appear to contribute to this problem:the uneven distribution of arrhythmia classes and the limited expressiveness of features learned by current models.To overcome these limitations,this study proposes a dual-path multimodal framework,termed DM-EHC(Dual-Path Multimodal ECG Heartbeat Classifier),for ECG-based heartbeat classification.The proposed framework links 1D ECG temporal features with 2D time–frequency features.By setting up the dual paths described above,the model can process more dimensions of feature information.The MIT-BIH arrhythmia database was selected as the baseline dataset for the experiments.Experimental results show that the proposed method outperforms single modalities and performs better for certain specific types of arrhythmias.The model achieved mean precision,recall,and F1 score of 95.14%,92.26%,and 93.65%,respectively.These results indicate that the framework is robust and has potential value in automated arrhythmia classification.展开更多
Near-Earth objects are important not only in studying the early formation of the Solar System,but also because they pose a serious hazard to humanity when they make close approaches to the Earth.Study of their physica...Near-Earth objects are important not only in studying the early formation of the Solar System,but also because they pose a serious hazard to humanity when they make close approaches to the Earth.Study of their physical properties can provide useful information on their origin,evolution,and hazard to human beings.However,it remains challenging to investigate small,newly discovered,near-Earth objects because of our limited observational window.This investigation seeks to determine the visible colors of near-Earth asteroids(NEAs),perform an initial taxonomic classification based on visible colors and analyze possible correlations between the distribution of taxonomic classification and asteroid size or orbital parameters.Observations were performed in the broadband BVRI Johnson−Cousins photometric system,applied to images from the Yaoan High Precision Telescope and the 1.88 m telescope at the Kottamia Astronomical Observatory.We present new photometric observations of 84 near-Earth asteroids,and classify 80 of them taxonomically,based on their photometric colors.We find that nearly half(46.3%)of the objects in our sample can be classified as S-complex,26.3%as C-complex,6%as D-complex,and 15.0%as X-complex;the remaining belong to the A-or V-types.Additionally,we identify three P-type NEAs in our sample,according to the Tholen scheme.The fractional abundances of the C/X-complex members with absolute magnitude H≥17.0 were more than twice as large as those with H<17.0.However,the fractions of C-and S-complex members with diameters≤1 km and>1 km are nearly equal,while X-complex members tend to have sub-kilometer diameters.In our sample,the C/D-complex objects are predominant among those with a Jovian Tisserand parameter of T_(J)<3.1.These bodies could have a cometary origin.C-and S-complex members account for a considerable proportion of the asteroids that are potentially hazardous.展开更多
Graph Neural Networks(GNNs)have proven highly effective for graph classification across diverse fields such as social networks,bioinformatics,and finance,due to their capability to learn complex graph structures.Howev...Graph Neural Networks(GNNs)have proven highly effective for graph classification across diverse fields such as social networks,bioinformatics,and finance,due to their capability to learn complex graph structures.However,despite their success,GNNs remain vulnerable to adversarial attacks that can significantly degrade their classification accuracy.Existing adversarial attack strategies primarily rely on label information to guide the attacks,which limits their applicability in scenarios where such information is scarce or unavailable.This paper introduces an innovative unsupervised attack method for graph classification,which operates without relying on label information,thereby enhancing its applicability in a broad range of scenarios.Specifically,our method first leverages a graph contrastive learning loss to learn high-quality graph embeddings by comparing different stochastic augmented views of the graphs.To effectively perturb the graphs,we then introduce an implicit estimator that measures the impact of various modifications on graph structures.The proposed strategy identifies and flips edges with the top-K highest scores,determined by the estimator,to maximize the degradation of the model’s performance.In addition,to defend against such attack,we propose a lightweight regularization-based defense mechanism that is specifically tailored to mitigate the structural perturbations introduced by our attack strategy.It enhances model robustness by enforcing embedding consistency and edge-level smoothness during training.We conduct experiments on six public TU graph classification datasets:NCI1,NCI109,Mutagenicity,ENZYMES,COLLAB,and DBLP_v1,to evaluate the effectiveness of our attack and defense strategies.Under an attack budget of 3,the maximum reduction in model accuracy reaches 6.67%on the Graph Convolutional Network(GCN)and 11.67%on the Graph Attention Network(GAT)across different datasets,indicating that our unsupervised method induces degradation comparable to state-of-the-art supervised attacks.Meanwhile,our defense achieves the highest accuracy recovery of 3.89%(GCN)and 5.00%(GAT),demonstrating improved robustness against structural perturbations.展开更多
Skin diseases affect millions worldwide.Early detection is key to preventing disfigurement,lifelong disability,or death.Dermoscopic images acquired in primary-care settings show high intra-class visual similarity and ...Skin diseases affect millions worldwide.Early detection is key to preventing disfigurement,lifelong disability,or death.Dermoscopic images acquired in primary-care settings show high intra-class visual similarity and severe class imbalance,and occasional imaging artifacts can create ambiguity for state-of-the-art convolutional neural networks(CNNs).We frame skin lesion recognition as graph-based reasoning and,to ensure fair evaluation and avoid data leakage,adopt a strict lesion-level partitioning strategy.Each image is first over-segmented using SLIC(Simple Linear Iterative Clustering)to produce perceptually homogeneous superpixels.These superpixels form the nodes of a region-adjacency graph whose edges encode spatial continuity.Node attributes are 1280-dimensional embeddings extracted with a lightweight yet expressive EfficientNet-B0 backbone,providing strong representational power at modest computational cost.The resulting graphs are processed by a five-layer Graph Attention Network(GAT)that learns to weight inter-node relationships dynamically and aggregates multi-hop context before classifying lesions into seven classes with a log-softmax output.Extensive experiments on the DermaMNIST benchmark show the proposed pipeline achieves 88.35%accuracy and 98.04%AUC,outperforming contemporary CNNs,AutoML approaches,and alternative graph neural networks.An ablation study indicates EfficientNet-B0 produces superior node descriptors compared with ResNet-18 and DenseNet,and that roughly five GAT layers strike a good balance between being too shallow and over-deep while avoiding oversmoothing.The method requires no data augmentation or external metadata,making it a drop-in upgrade for clinical computer-aided diagnosis systems.展开更多
With the recent increase in data volume and diversity,traditional text representation techniques are struggling to capture context,particularly in environments with sparse data.To address these challenges,this study p...With the recent increase in data volume and diversity,traditional text representation techniques are struggling to capture context,particularly in environments with sparse data.To address these challenges,this study proposes a new model,the Masked Joint Representation Model(MJRM).MJRM approximates the original hypothesis by leveraging multiple elements in a limited context.It dynamically adapts to changes in characteristics based on data distribution through three main components.First,masking-based representation learning,termed selective dynamic masking,integrates topic modeling and sentiment clustering to generate and train multiple instances across different data subsets,whose predictions are then aggregated with optimized weights.This design alleviates sparsity,suppresses noise,and preserves contextual structures.Second,regularization-based improvements are applied.Third,techniques for addressing sparse data are used to perform final inference.As a result,MJRM improves performance by up to 4%compared to existing AI techniques.In our experiments,we analyzed the contribution of each factor,demonstrating that masking,dynamic learning,and aggregating multiple instances complement each other to improve performance.This demonstrates that a masking-based multi-learning strategy is effective for context-aware sparse text classification,and can be useful even in challenging situations such as data shortage or data distribution variations.We expect that the approach can be extended to diverse fields such as sentiment analysis,spam filtering,and domain-specific document classification.展开更多
Container transportation is pivotal in global trade due to its efficiency,safety,and cost-effectiveness.However,structural defects—particularly in grapple slots—can result in cargo damage,financial loss,and elevated...Container transportation is pivotal in global trade due to its efficiency,safety,and cost-effectiveness.However,structural defects—particularly in grapple slots—can result in cargo damage,financial loss,and elevated safety risks,including container drops during lifting operations.Timely and accurate inspection before and after transit is therefore essential.Traditional inspection methods rely heavily on manual observation of internal and external surfaces,which are time-consuming,resource-intensive,and prone to subjective errors.Container roofs pose additional challenges due to limited visibility,while grapple slots are especially vulnerable to wear from frequent use.This study proposes a two-stage automated detection framework targeting defects in container roof grapple slots.In the first stage,YOLOv7 is employed to localize grapple slot regions with high precision.In the second stage,ResNet50 classifies the extracted slots as either intact or defective.The results from both stages are integrated into a human-machine interface for real-time visualization and user verification.Experimental evaluations demonstrate that YOLOv7 achieves a 99%detection rate at 100 frames per second(FPS),while ResNet50 attains 87%classification accuracy at 34 FPS.Compared to some state of the arts,the proposed system offers significant speed,reliability,and usability improvements,enabling efficient defect identification and visual reconfirmation via the interface.展开更多
Evaluating the adversarial robustness of classification algorithms in machine learning is a crucial domain.However,current methods lack measurable and interpretable metrics.To address this issue,this paper introduces ...Evaluating the adversarial robustness of classification algorithms in machine learning is a crucial domain.However,current methods lack measurable and interpretable metrics.To address this issue,this paper introduces a visual evaluation index named confidence centroid skewing quadrilateral,which is based on a classification confidence-based confusion matrix,offering a quantitative and visual comparison of the adversarial robustness among different classification algorithms,and enhances intuitiveness and interpretability of attack impacts.We first conduct a validity test and sensitive analysis of the method.Then,prove its effectiveness through the experiments of five classification algorithms including artificial neural network(ANN),logistic regression(LR),support vector machine(SVM),convolutional neural network(CNN)and transformer against three adversarial attacks such as fast gradient sign method(FGSM),DeepFool,and projected gradient descent(PGD)attack.展开更多
The deep learning technology has shown impressive performance in various vision tasks such as image classification, object detection and semantic segmentation. In particular, recent advances of deep learning technique...The deep learning technology has shown impressive performance in various vision tasks such as image classification, object detection and semantic segmentation. In particular, recent advances of deep learning techniques bring encouraging performance to fine-grained image classification which aims to distinguish subordinate-level categories, such as bird species or dog breeds. This task is extremely challenging due to high intra-class and low inter-class variance. In this paper, we review four types of deep learning based fine-grained image classification approaches, including the general convolutional neural networks （CNNs）, part detection based, ensemble of networks based and visual attention based fine-grained image classification approaches. Besides, the deep learning based semantic segmentation approaches are also covered in this paper. The region proposal based and fully convolutional networks based approaches for semantic segmentation are introduced respectively.展开更多
文摘The remote sensing ships’fine-grained classification technology makes it possible to identify certain ship types in remote sensing images,and it has broad application prospects in civil and military fields.However,the current model does not examine the properties of ship targets in remote sensing images with mixed multi-granularity features and a complicated backdrop.There is still an opportunity for future enhancement of the classification impact.To solve the challenges brought by the above characteristics,this paper proposes a Metaformer and Residual fusion network based on Visual Attention Network(VAN-MR)for fine-grained classification tasks.For the complex background of remote sensing images,the VAN-MR model adopts the parallel structure of large kernel attention and spatial attention to enhance the model’s feature extraction ability of interest targets and improve the classification performance of remote sensing ship targets.For the problem of multi-grained feature mixing in remote sensing images,the VAN-MR model uses a Metaformer structure and a parallel network of residual modules to extract ship features.The parallel network has different depths,considering both high-level and lowlevel semantic information.The model achieves better classification performance in remote sensing ship images with multi-granularity mixing.Finally,the model achieves 88.73%and 94.56%accuracy on the public fine-grained ship collection-23(FGSC-23)and FGSCR-42 datasets,respectively,while the parameter size is only 53.47 M,the floating point operations is 9.9 G.The experimental results show that the classification effect of VAN-MR is superior to that of traditional CNNs model and visual model with Transformer structure under the same parameter quantity.
基金supported by National Basic Research Program of China (973 Program) (No. 2015CB352502)National Nature Science Foundation of China (No. 61573026)Beijing Nature Science Foundation (No. L172037)
文摘Fine-grained image classification, which aims to distinguish images with subtle distinctions, is a challenging task for two main reasons: lack of sufficient training data for every class and difficulty in learning discriminative features for representation. In this paper, to address the two issues, we propose a two-phase framework for recognizing images from unseen fine-grained classes, i.e., zeroshot fine-grained classification. In the first feature learning phase, we finetune deep convolutional neural networks using hierarchical semantic structure among fine-grained classes to extract discriminative deep visual features. Meanwhile, a domain adaptation structure is induced into deep convolutional neural networks to avoid domain shift from training data to test data. In the second label inference phase, a semantic directed graph is constructed over attributes of fine-grained classes. Based on this graph, we develop a label propagation algorithm to infer the labels of images in the unseen classes. Experimental results on two benchmark datasets demonstrate that our model outperforms the state-of-the-art zero-shot learning models. In addition, the features obtained by our feature learning model also yield significant gains when they are used by other zero-shot learning models, which shows the flexility of our model in zero-shot finegrained classification.
文摘With the rapid development of the Internet of things and e-commerce, feature-based image retrieval and classification have become a serious challenge for shoppers searching websites for relevant product information. The last decade has witnessed great interest in research on content-based feature extraction techniques. Moreover, semantic attributes cannot fully express the rich image information. This paper designs and trains a deep convolutional neural network that the convolution kernel size and the order of network connection are based on the high efficiency of the filter capacity and coverage. To solve the problem of long training time and high resource share of deep convolutional neural network, this paper designed a shallow convolutional neural network to achieve the similar classification accuracy. The deep and shallow convolutional neural networks have data pre-processing, feature extraction and softmax classification. To evaluate the classification performance of the network, experiments were conducted using a public database Caltech256 and a homemade product image database containing 15 species of garment and 5 species of shoes on a total of 20,000 color images from shopping websites. Compared with the classification accuracy of combining content-based feature extraction techniques with traditional support vector machine techniques from 76.3% to 86.2%, the deep convolutional neural network obtains an impressive state-of-the-art classification accuracy of 92.1%, and the shallow convolutional neural network reached a classification accuracy of 90.6%. Moreover, the proposed convolutional neural networks can be integrated and implemented in other colour image database.
基金supported by the National Natural Science Foundation of China(42030102,42371321).
文摘Accurate fine-grained geospatial scene classification using remote sensing imagery is essential for a wide range of applications.However,existing approaches often rely on manually zooming remote sensing images at different scales to create typical scene samples.This approach fails to adequately support the fixed-resolution image interpretation requirements in real-world scenarios.To address this limitation,we introduce the million-scale fine-grained geospatial scene classification dataset(MEET),which contains over 1.03 million zoom-free remote sensing scene samples,manually annotated into 80 fine-grained categories.In MEET,each scene sample follows a scene-in-scene layout,where the central scene serves as the reference,and auxiliary scenes provide crucial spatial context for fine-grained classification.Moreover,to tackle the emerging challenge of scene-in-scene classification,we present the context-aware transformer(CAT),a model specifically designed for this task,which adaptively fuses spatial context to accurately classify the scene samples.CAT adaptively fuses spatial context to accurately classify the scene samples by learning attentional features that capture the relationships between the center and auxiliary scenes.Based on MEET,we establish a comprehensive benchmark for fine-grained geospatial scene classification,evaluating CAT against 11 competitive baselines.The results demonstrate that CAT significantly outperforms these baselines,achieving a 1.88%higher balanced accuracy(BA)with the Swin-Large backbone,and a notable 7.87%improvement with the Swin-Huge backbone.Further experiments validate the effectiveness of each module in CAT and show the practical applicability of CAT in the urban functional zone mapping.The source code and dataset will be publicly available at https://jerrywyn.github.io/project/MEET.html.
基金Supported by the National Natural Science Foundation of China(61601176)。
文摘In this paper,we propose hierarchical attention dual network(DNet)for fine-grained image classification.The DNet can randomly select pairs of inputs from the dataset and compare the differences between them through hierarchical attention feature learning,which are used simultaneously to remove noise and retain salient features.In the loss function,it considers the losses of difference in paired images according to the intra-variance and inter-variance.In addition,we also collect the disaster scene dataset from remote sensing images and apply the proposed method to disaster scene classification,which contains complex scenes and multiple types of disasters.Compared to other methods,experimental results show that the DNet with hierarchical attention is robust to different datasets and performs better.
基金supported by the Beijing Natural Science Foundation(No.5252014)the Open Fund of The Key Laboratory of Urban Ecological Environment Simulation and Protection,Ministry of Ecology and Environment of the People's Republic of China (No.UEESP-202502)the National Natural Science Foundation of China (No.62303063&32371874)。
文摘Bird monitoring and protection are essential for maintaining biodiversity,and fine-grained bird classification has become a key focus in this field.Audio-visual modalities provide critical cues for this task,but robust feature extraction and efficient fusion remain major challenges.We introduce a multi-stage fine-grained audiovisual fusion network(MSFG-AVFNet) for fine-grained bird species classification,which addresses these challenges through two key components:(1) the audiovisual feature extraction module,which adopts a multi-stage finetuning strategy to provide high-quality unimodal features,laying a solid foundation for modality fusion;(2) the audiovisual feature fusion module,which combines a max pooling aggregation strategy with a novel audiovisual loss function to achieve effective and robust feature fusion.Experiments were conducted on the self-built AVB81and the publicly available SSW60 datasets,which contain data from 81 and 60 bird species,respectively.Comprehensive experiments demonstrate that our approach achieves notable performance gains,outperforming existing state-of-the-art methods.These results highlight its effectiveness in leveraging audiovisual modalities for fine-grained bird classification and its potential to support ecological monitoring and biodiversity research.
基金supported by National Key Research and Development Program of China(2022YFB3104903)S&T Program of Hebei(No.SZX2020034).
文摘Intelligent vehicle applications provide convenience but raise privacy and security concerns.Misuse of sensitive data,including vehicle location,and facial recognition information,poses a threat to user privacy.Hence,traffic classification is vital for promptly overseeing and controlling applications with sensitive information.In this paper,we propose ETNet,a framework that combines multiple features and leverages self-attention mechanisms to learn deep relationships between packets.ET-Net employs a multisimilarity triplet network to extract features from raw bytes,and exploits self-attention to capture long-range dependencies within packets in a session and contextual information features.Additionally,we utilizing the loss function to more effectively integrate information acquired from both byte sequences and their corresponding lengths.Through simulated evaluations on datasets with similar attributes,ET-Net demonstrates the ability to finely distinguish between nine categories of applications,achieving superior results compared to existing methods.
基金the National Natural Science Foundation of China(Project Nos.61521002 and 61772298)a Research Grant of Beijing Higher Institution Engineering Research CenterTsinghua–Tencent Joint Laboratory for Internet Innovation Technology。
文摘In this paper, we introduce an image dataset for fine-grained classification of dog breeds: the Tsinghua Dogs Dataset. It is currently the largest dataset for fine-grained classification of dogs, including 130 dog breeds and 70,428 real-world images. It has only one dog in each image and provides annotated bounding boxes for the whole body and head. In comparison to previous similar datasets, it contains more breeds and more carefully chosen images for each breed. The diversity within each breed is greater,with between 200 and 7000+ images for each breed.Annotation of the whole body and head makes the dataset not only suitable for the improvement of finegrained image classification models based on overall features, but also for those locating local informative parts. We show that dataset provides a tough challenge by benchmarking several state-of-the-art deep neural models. The dataset is available for academic purposes at https://cg.cs.tsinghua.edu.cn/ThuDogs/.
基金This work was financially supported by the National Key Research and Development Project(Grant No.2020YFD1100601)。
文摘The value of grape cultivars varies.The use of a mixture of cultivars can negate the benefits of improved cultivars and hamper the protection of genetic resources and the identification of new hybrid cultivars.Classifying cultivars based on their leaves is therefore highly practical.Transplanted grape seedlings take years to bear fruit,but leaves mature in months.Foliar morphology differs among cultivars,so identifying cultivars based on leaves is feasible.Different cultivars,however,can be bred from the same parents,so the leaves of some cultivars can have similar morphologies.In this work,a pyramid residual convolution neural network was developed to classify images of eleven grape cultivars.The model extracts multi-scale feature maps of the leaf images through the convolution layer and enters them into three residual convolution neural networks.Features are fused by adding the value of the convolution kernel feature matrix to enhance the attention on the edge and center regions of the leaves and classify the images.The results indicated that the average accuracy of the model was 92.26%for the proposed leaf dataset.The proposed model is superior to previous models and provides a reliable method for the fine-grained classification and identification of plant cultivars.
文摘In recent decades,the proliferation of email communication has markedly escalated,resulting in a concomitant surge in spam emails that congest networks and presenting security risks.This study introduces an innovative spam detection method utilizing the Horse Herd Optimization Algorithm(HHOA),designed for binary classification within multi⁃objective framework.The method proficiently identifies essential features,minimizing redundancy and improving classification precision.The suggested HHOA attained an impressive accuracy of 97.21%on the Kaggle email dataset,with precision of 94.30%,recall of 90.50%,and F1⁃score of 92.80%.Compared to conventional techniques,such as Support Vector Machine(93.89%accuracy),Random Forest(96.14%accuracy),and K⁃Nearest Neighbours(92.08%accuracy),HHOA exhibited enhanced performance with reduced computing complexity.The suggested method demonstrated enhanced feature selection efficiency,decreasing the number of selected features while maintaining high classification accuracy.The results underscore the efficacy of HHOA in spam identification and indicate its potential for further applications in practical email filtering systems.
基金supported and funded by the Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University(IMSIU)(grant number IMSIU-DDRSP2601).
文摘Visual diagnosis of skin cancer is challenging due to subtle inter-class similarities,variations in skin texture,the presence of hair,and inconsistent illumination.Deep learning models have shown promise in assisting early detection,yet their performance is often limited by the severe class imbalance present in dermoscopic datasets.This paper proposes CANNSkin,a skin cancer classification framework that integrates a convolutional autoencoder with latent-space oversampling to address this imbalance.The autoencoder is trained to reconstruct lesion images,and its latent embeddings are used as features for classification.To enhance minority-class representation,the Synthetic Minority Oversampling Technique(SMOTE)is applied directly to the latent vectors before classifier training.The encoder and classifier are first trained independently and later fine-tuned end-to-end.On the HAM10000 dataset,CANNSkin achieves an accuracy of 93.01%,a macro-F1 of 88.54%,and an ROC–AUC of 98.44%,demonstrating strong robustness across ten test subsets.Evaluation on the more complex ISIC 2019 dataset further confirms the model’s effectiveness,where CANNSkin achieves 94.27%accuracy,93.95%precision,94.09%recall,and 99.02%F1-score,supported by high reconstruction fidelity(PSNR 35.03 dB,SSIM 0.86).These results demonstrate the effectiveness of our proposed latent-space balancing and fine-tuned representation learning as a new benchmark method for robust and accurate skin cancer classification across heterogeneous datasets.
基金supported,in part,by the National Nature Science Foundation of China under Grant 62272236,62376128 and 62306139the Natural Science Foundation of Jiangsu Province under Grant BK20201136,BK20191401.
文摘Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in complex backgrounds,small target objects,and limited training data,leading to poor recognition.Fine-grained images exhibit“small inter-class differences,”and while second-order feature encoding enhances discrimination,it often requires dual Convolutional Neural Networks(CNN),increasing training time and complexity.This study proposes a model integrating discriminative region localization and efficient second-order feature encoding.By ranking feature map channels via a fully connected layer,it selects high-importance channels to generate an enhanced map,accurately locating discriminative regions.Cropping and erasing augmentations further refine recognition.To improve efficiency,a novel second-order feature encoding module generates an attention map from the fourth convolutional group of Residual Network 50 layers(ResNet-50)and multiplies it with features from the fifth group,producing second-order features while reducing dimensionality and training time.Experiments on Caltech-University of California,San Diego Birds-200-2011(CUB-200-2011),Stanford Car,and Fine-Grained Visual Classification of Aircraft(FGVC Aircraft)datasets show state-of-the-art accuracy of 88.9%,94.7%,and 93.3%,respectively.
基金supported by the Innovative Human Resource Development for Local Intel-lectualization program through the Institute of Information&Communications Technology Planning&Evaluation(IITP)grant funded by the Korea government(MSIT)(No.IITP-2026-2020-0-01741)the research fund of Hanyang University(HY-2025-1110).
文摘Arrhythmias are a frequently occurring phenomenon in clinical practice,but how to accurately dis-tinguish subtle rhythm abnormalities remains an ongoing difficulty faced by the entire research community when conducting ECG-based studies.From a review of existing studies,two main factors appear to contribute to this problem:the uneven distribution of arrhythmia classes and the limited expressiveness of features learned by current models.To overcome these limitations,this study proposes a dual-path multimodal framework,termed DM-EHC(Dual-Path Multimodal ECG Heartbeat Classifier),for ECG-based heartbeat classification.The proposed framework links 1D ECG temporal features with 2D time–frequency features.By setting up the dual paths described above,the model can process more dimensions of feature information.The MIT-BIH arrhythmia database was selected as the baseline dataset for the experiments.Experimental results show that the proposed method outperforms single modalities and performs better for certain specific types of arrhythmias.The model achieved mean precision,recall,and F1 score of 95.14%,92.26%,and 93.65%,respectively.These results indicate that the framework is robust and has potential value in automated arrhythmia classification.
基金funded by the China National Space Administration(KJSP2023020105)supported by the National Key R&D Program of China(Grant No.2023YFA1608100)+2 种基金the NSFC(Grant No.62227901)the Minor Planet Foundationsupported by the Egyptian Science,Technology&Innovation Funding Authority(STDF)under Grant No.48102.
文摘Near-Earth objects are important not only in studying the early formation of the Solar System,but also because they pose a serious hazard to humanity when they make close approaches to the Earth.Study of their physical properties can provide useful information on their origin,evolution,and hazard to human beings.However,it remains challenging to investigate small,newly discovered,near-Earth objects because of our limited observational window.This investigation seeks to determine the visible colors of near-Earth asteroids(NEAs),perform an initial taxonomic classification based on visible colors and analyze possible correlations between the distribution of taxonomic classification and asteroid size or orbital parameters.Observations were performed in the broadband BVRI Johnson−Cousins photometric system,applied to images from the Yaoan High Precision Telescope and the 1.88 m telescope at the Kottamia Astronomical Observatory.We present new photometric observations of 84 near-Earth asteroids,and classify 80 of them taxonomically,based on their photometric colors.We find that nearly half(46.3%)of the objects in our sample can be classified as S-complex,26.3%as C-complex,6%as D-complex,and 15.0%as X-complex;the remaining belong to the A-or V-types.Additionally,we identify three P-type NEAs in our sample,according to the Tholen scheme.The fractional abundances of the C/X-complex members with absolute magnitude H≥17.0 were more than twice as large as those with H<17.0.However,the fractions of C-and S-complex members with diameters≤1 km and>1 km are nearly equal,while X-complex members tend to have sub-kilometer diameters.In our sample,the C/D-complex objects are predominant among those with a Jovian Tisserand parameter of T_(J)<3.1.These bodies could have a cometary origin.C-and S-complex members account for a considerable proportion of the asteroids that are potentially hazardous.
基金funded by the National Key Research and Development Program of China(Grant No.2024YFE0209000)the NSFC(Grant No.U23B2019).
文摘Graph Neural Networks(GNNs)have proven highly effective for graph classification across diverse fields such as social networks,bioinformatics,and finance,due to their capability to learn complex graph structures.However,despite their success,GNNs remain vulnerable to adversarial attacks that can significantly degrade their classification accuracy.Existing adversarial attack strategies primarily rely on label information to guide the attacks,which limits their applicability in scenarios where such information is scarce or unavailable.This paper introduces an innovative unsupervised attack method for graph classification,which operates without relying on label information,thereby enhancing its applicability in a broad range of scenarios.Specifically,our method first leverages a graph contrastive learning loss to learn high-quality graph embeddings by comparing different stochastic augmented views of the graphs.To effectively perturb the graphs,we then introduce an implicit estimator that measures the impact of various modifications on graph structures.The proposed strategy identifies and flips edges with the top-K highest scores,determined by the estimator,to maximize the degradation of the model’s performance.In addition,to defend against such attack,we propose a lightweight regularization-based defense mechanism that is specifically tailored to mitigate the structural perturbations introduced by our attack strategy.It enhances model robustness by enforcing embedding consistency and edge-level smoothness during training.We conduct experiments on six public TU graph classification datasets:NCI1,NCI109,Mutagenicity,ENZYMES,COLLAB,and DBLP_v1,to evaluate the effectiveness of our attack and defense strategies.Under an attack budget of 3,the maximum reduction in model accuracy reaches 6.67%on the Graph Convolutional Network(GCN)and 11.67%on the Graph Attention Network(GAT)across different datasets,indicating that our unsupervised method induces degradation comparable to state-of-the-art supervised attacks.Meanwhile,our defense achieves the highest accuracy recovery of 3.89%(GCN)and 5.00%(GAT),demonstrating improved robustness against structural perturbations.
基金funded by the Deanship of Graduate Studies and Scientific Research at Jouf University under grant No.(DGSSR-2025-02-01296).
文摘Skin diseases affect millions worldwide.Early detection is key to preventing disfigurement,lifelong disability,or death.Dermoscopic images acquired in primary-care settings show high intra-class visual similarity and severe class imbalance,and occasional imaging artifacts can create ambiguity for state-of-the-art convolutional neural networks(CNNs).We frame skin lesion recognition as graph-based reasoning and,to ensure fair evaluation and avoid data leakage,adopt a strict lesion-level partitioning strategy.Each image is first over-segmented using SLIC(Simple Linear Iterative Clustering)to produce perceptually homogeneous superpixels.These superpixels form the nodes of a region-adjacency graph whose edges encode spatial continuity.Node attributes are 1280-dimensional embeddings extracted with a lightweight yet expressive EfficientNet-B0 backbone,providing strong representational power at modest computational cost.The resulting graphs are processed by a five-layer Graph Attention Network(GAT)that learns to weight inter-node relationships dynamically and aggregates multi-hop context before classifying lesions into seven classes with a log-softmax output.Extensive experiments on the DermaMNIST benchmark show the proposed pipeline achieves 88.35%accuracy and 98.04%AUC,outperforming contemporary CNNs,AutoML approaches,and alternative graph neural networks.An ablation study indicates EfficientNet-B0 produces superior node descriptors compared with ResNet-18 and DenseNet,and that roughly five GAT layers strike a good balance between being too shallow and over-deep while avoiding oversmoothing.The method requires no data augmentation or external metadata,making it a drop-in upgrade for clinical computer-aided diagnosis systems.
基金supported by the SungKyunKwan University and the BK21 FOUR(Graduate School Innovation)funded by the Ministry of Education(MOE,Korea)and National Research Foundation of Korea(NRF).
文摘With the recent increase in data volume and diversity,traditional text representation techniques are struggling to capture context,particularly in environments with sparse data.To address these challenges,this study proposes a new model,the Masked Joint Representation Model(MJRM).MJRM approximates the original hypothesis by leveraging multiple elements in a limited context.It dynamically adapts to changes in characteristics based on data distribution through three main components.First,masking-based representation learning,termed selective dynamic masking,integrates topic modeling and sentiment clustering to generate and train multiple instances across different data subsets,whose predictions are then aggregated with optimized weights.This design alleviates sparsity,suppresses noise,and preserves contextual structures.Second,regularization-based improvements are applied.Third,techniques for addressing sparse data are used to perform final inference.As a result,MJRM improves performance by up to 4%compared to existing AI techniques.In our experiments,we analyzed the contribution of each factor,demonstrating that masking,dynamic learning,and aggregating multiple instances complement each other to improve performance.This demonstrates that a masking-based multi-learning strategy is effective for context-aware sparse text classification,and can be useful even in challenging situations such as data shortage or data distribution variations.We expect that the approach can be extended to diverse fields such as sentiment analysis,spam filtering,and domain-specific document classification.
文摘Container transportation is pivotal in global trade due to its efficiency,safety,and cost-effectiveness.However,structural defects—particularly in grapple slots—can result in cargo damage,financial loss,and elevated safety risks,including container drops during lifting operations.Timely and accurate inspection before and after transit is therefore essential.Traditional inspection methods rely heavily on manual observation of internal and external surfaces,which are time-consuming,resource-intensive,and prone to subjective errors.Container roofs pose additional challenges due to limited visibility,while grapple slots are especially vulnerable to wear from frequent use.This study proposes a two-stage automated detection framework targeting defects in container roof grapple slots.In the first stage,YOLOv7 is employed to localize grapple slot regions with high precision.In the second stage,ResNet50 classifies the extracted slots as either intact or defective.The results from both stages are integrated into a human-machine interface for real-time visualization and user verification.Experimental evaluations demonstrate that YOLOv7 achieves a 99%detection rate at 100 frames per second(FPS),while ResNet50 attains 87%classification accuracy at 34 FPS.Compared to some state of the arts,the proposed system offers significant speed,reliability,and usability improvements,enabling efficient defect identification and visual reconfirmation via the interface.
文摘Evaluating the adversarial robustness of classification algorithms in machine learning is a crucial domain.However,current methods lack measurable and interpretable metrics.To address this issue,this paper introduces a visual evaluation index named confidence centroid skewing quadrilateral,which is based on a classification confidence-based confusion matrix,offering a quantitative and visual comparison of the adversarial robustness among different classification algorithms,and enhances intuitiveness and interpretability of attack impacts.We first conduct a validity test and sensitive analysis of the method.Then,prove its effectiveness through the experiments of five classification algorithms including artificial neural network(ANN),logistic regression(LR),support vector machine(SVM),convolutional neural network(CNN)and transformer against three adversarial attacks such as fast gradient sign method(FGSM),DeepFool,and projected gradient descent(PGD)attack.
基金supported by the National Natural Science Foundation of China(Nos.61373121 and 61328205)Program for Sichuan Provincial Science Fund for Distinguished Young Scholars(No.13QNJJ0149)+1 种基金the Fundamental Research Funds for the Central UniversitiesChina Scholarship Council(No.201507000032)
文摘The deep learning technology has shown impressive performance in various vision tasks such as image classification, object detection and semantic segmentation. In particular, recent advances of deep learning techniques bring encouraging performance to fine-grained image classification which aims to distinguish subordinate-level categories, such as bird species or dog breeds. This task is extremely challenging due to high intra-class and low inter-class variance. In this paper, we review four types of deep learning based fine-grained image classification approaches, including the general convolutional neural networks （CNNs）, part detection based, ensemble of networks based and visual attention based fine-grained image classification approaches. Besides, the deep learning based semantic segmentation approaches are also covered in this paper. The region proposal based and fully convolutional networks based approaches for semantic segmentation are introduced respectively.
