Applied Intelligence最新文献

Attribute reduction can improve the information processing efficiency by removing redundant attributes without degrading the information in the dataset. The delineation of information granularity and fault tolerance in attribute reduction are essential. To address the above issues, an attribute reduction algorithm based on variable precision neighborhood rough set with bivariate and inclusion degree is proposed. Firstly, the concepts of bivariate and inclusion degree are introduced into the variable precision neighborhood rough set. The bivariate is used to compute the distance between the information granules and thus divide the information granules. The inclusion degree is used to increase the fault tolerance of the model when calculating the upper and lower approximations. Secondly, novel model enhances ability to divide information granules and fault tolerance. On this foundation, an algorithm for attribute reduction is proposed. Finally, this paper ensures the validity of the experiments by conducting them on real datasets. The experimental results show that the algorithm is able to reduce the redundant attributes in the dataset.

Melanoma skin cancer is an important, growing, and among the most lethal skin cancer forms, but could easily be cured if detected at an earlier stage. The primary cause of skin cancer is the abnormal growth of skin cells, often resulting from prolonged exposure to sunlight. Unfortunately, identifying malignant skin development at an early stage is both costly and challenging. The grading of skin cancer is determined based on the location of the tumor and the type of cells involved. Accurate classification of lesions requires a particular range of precision and recall, presenting a significant challenge in dermatology. This paper proposes a new system, the Composite Convolutional Neural Network (CCNN) skin cancer detector (SCD), based on the use of more sophisticated deep learning systems known as composite convolutional neural networks. The framework is applied in three primary stages. During the first step, noise, normalization, and increment of dermoscopic images are performed to improve the quality of data and variability. The next stage will utilize transfer learning using the models ResNet50, InceptionV3, and VGG16 to identify the high-level discriminative features. In the classification module, skin lesion classification is done using a customized CNN where the feature dimensionality reduction is done using an attention-based autoencoder to enhance the generalization of the model. Lastly, the deep features obtained are tested with convolutional machine learning classifiers, such as K - nearest neighbor (KNN), random forest classifier (RFC), and support-vector machine (SVM), which allow to compare the performance of the derived deep features; the accuracy of 93.53%, the precision of 94.02%, and the recall of 93.61% were achieved, which leads to the conclusion of the strength of the proposed approach and its possible use in clinical dermatologic diagnosis.

Gait recognition technology represents a promising biometric identification method with extensive applications in fields such as visual surveillance. However, addressing the few-shot learning challenge remains an urgent issue in practical applications. In this paper, we propose a novel Dual-Scale Spatial-Temporal Aggregation Network (DSAN), which extracts gait features at different temporal scales through a dual-branch architecture. The Multi-Scale Temporal Aggregation Module (MSTA) is designed to aggregate multi-scale local temporal information from long-term gait features, employing an attention mechanism to reduce redundant information and enhance temporal feature representation. Additionally, the Covariance-guided Spatial-Temporal Attention Module (CoSTA) integrates short-term features into long-term representations through global covariance pooling and attention mechanisms, further improving feature expressiveness. Furthermore, we introduce the View Variation Adversarial Normalization (VVAN), which mitigates cross-view discrepancies under few-shot conditions via adversarial learning. Different from GaitGL, DSAN explicitly captures long-term and short-term dynamics and performs cross-scale fusion via MSTA and CoSTA. In contrast to GaitDAN’s multi-discriminator formulation, VVAN employs a single multi-class view discriminator, requiring only a change in the classifier output dimension when the view set varies. Extensive experiments are conducted on the CASIA-B and OUMVLP datasets, with DSAN evaluated in a one-shot setting on CASIA-B. The experimental results demonstrate that DSAN achieves state-of-the-art performance in gait recognition tasks and effectively alleviates the adverse impact of the one-shot problem on recognition accuracy.

Multi-source information fusion is a key technique in big data technology, essential for data mining and knowledge discovery. When merging traditional multi-source systems into single systems, data loss is a common issue. Multi-granularity information fusion methods are designed to solve this problem. However, there is limited research on how to use these methods dynamically to extract valuable information when there are changes in three dimensions: objects, information sources, and the number of attributes. Additionally, information fusion in multi-source fuzzy information systems can better handle uncertainties in data. This paper proposes a method for calculating multi-granularity fusion operators based on a similarity relationship matrix and creates an incremental update fusion mechanism for six different scenarios, such as adding or removing objects, information sources, and attributes. Experimental results on 12 public datasets show that our proposed dynamic fusion method has clear advantages in dealing with complex changes and can update the fusion operator more efficiently.

Deep neural networks are among the most successful algorithms in terms of performance and scalability across different domains. However, since these networks are black boxes, their usability is severely restricted due to a lack of interpretability. Existing interpretability methods do not address the analysis of time-series-based networks specifically enough. This paper shows that an analysis in the frequency domain can not only highlight relevant areas in the input signal better than existing methods but is also more robust to fluctuations in the signal. In this paper, FreqAtt is presented — a framework that enables post-hoc interpretation of time-series analysis. To achieve this, the relevant frequencies are evaluated, and the signal is either filtered or the relevant input data is marked. FreqAtt is evaluated using a wide range of statistical metrics to provide a broad overview of its performance. The results show that using frequency-based attribution, especially in combination with traditional attribution on top of the frequency-optimized signal, provides strong performance across different metrics.

Tracking a fixed-wing unmanned aerial vehicle (UAV) plays a key role in the aerospace applications, but this research area lacks a high-diversity and large-scale benchmark dataset, which is a key factor for both the comprehensive evaluation of UAV to UAV (UAV2UAV) tracking algorithms and the development of deep learning methods. It is thus essential to construct a UAV2UAV dataset to fill the gap and advance the research in this domain. To this end, we release here a large-scale dataset, UAV2UAV-2 dataset, for a fixed-wing UAV tracking. Our UAV2UAV-2 dataset includes 49 videos with more than 27k frames in total. A training dataset with more than 4k images and various background scenarios is also provided, which can be used to further validate the training process. A comprehensive performance evaluation of 24 tracking methods is carried out on this dataset. Three representative tracking methods which are trained on the training dataset achieve superior results compared to the original corresponding methods. This indicates that training with dedicated data can improve tracking performance, and this approach can be extended to more tracking algorithms. A detailed analysis of the performance results is also provided. The data will be available at: https://github.com/zxysysu/UAV2UAV-2.

Robotic task planning in open-world environments remains challenging due to the requirements of cross-modal semantic understanding, action feasibility verification, and reliable closed-loop correction. This paper presents TRTP, a three-stage robust task-planning framework that combines vision-language models with digital twin simulation to address these issues. The framework consists of three tightly integrated components. The spatial prompt generation module converts scene images into structured spatial prompts that encode relations such as orientation and containment. The task planning module uses these prompts, along with task instructions and demonstration videos, to produce an initial sequence of executable actions. The digital twin simulation module evaluates spatial reachability and physical feasibility in a virtual environment and provides structured corrective feedback. Together, these modules form a complete multimodal, closed-loop planning pipeline that mitigates failures arising from incomplete spatial modeling or the lack of physical validation. We further develop a multi-model inference and cross-scoring data generation pipeline to support fine-tuning of the vision-language models used in the spatial-prompt and task planning modules, thereby reducing reliance on manual annotation. Experimental evaluations demonstrate that TRTP achieves marked improvements in task consistency, physical feasibility, and overall success rate compared with baselines that omit spatial prompts or feedback. The results confirm the complementary contributions of the three stages and show that TRTP generalizes effectively to diverse real-world robotic scenarios. Home page: https://trtp.github.io/

Zero-shot object navigation involves locating target objects in unseen environments, a task fundamental to embodied intelligence. Traditional approaches, particularly recent vision-language navigation models, leverage Large Language Models (LLMs) to enable multimodal reasoning based on real-time visual perception. However, two key challenges remain: (1) Semantic mismatches emerge between the global map and the environment when frontier selection optimizes a singular optimization criterion; and (2) General-purpose LLMs are challenging to optimize for navigation tasks, leading to limited navigation-specific reasoning capabilities. To address these issues, we propose Benefit Frontier Semantic Map (BFSMap) to enable human-like semantic exploration through iterative reasoning. First, BFSMap reformulates semantic mapping as an image captioning task, integrating multiple maps to derive an optimal frontier that balances benefit and cost, thereby alleviating vision–language misalignment in prior end-to-end methods. Second, we introduce a lightweight semantic-aware benefit prediction module (LightSA), trained from scratch using a novel prompt learning strategy for cross-view semantic reasoning to update the semantic maps. Third, we design a modular object-aware decision-making policy that mimics human-like reasoning to identify the target object and correct suboptimal paths promptly. Our model achieves state-of-the-art performance (+2.8% SR on HM3D, +1.2% SR on MP3D compared to baseline) without relying on LLMs, demonstrating the promise of efficient navigation models based on human-like reasoning for unknown environment. Our code will be available.

With the rapid advancement of remote sensing (RS) technology, the difficulty of acquiring multi-source RS data is significantly decreasing. Land use/land cover (LULC) classification methods that fuse hyperspectral (HS) image and light detection and ranging (LiDAR) have become a current research hotspot. However, current mainstream methods primarily focus on extracting the most salient features from HS image and LiDAR data. They emphasize the rich spectral information of HS image and the precise elevation information of LiDAR, often overlooking the important frequency-domain characteristics inherent in both data sources. Based on this, we propose a novel network framework called frequency coupled fusion convolutional neural network (FCFCNN), which aims to enhance classification performance by integrating frequency features from HS image and LiDAR data as supplementary information in the feature fusion process. The network is designed with a dual-branch structure. The first branch focuses on frequency information, employing a frequency-splitting module to separately extract high-frequency and low-frequency features from HS image and LiDAR data. Subsequently, the local enhanced position attention module (LEPAM) and coupling strategy are used to fuse these high-frequency and low-frequency features, respectively. The second branch focuses on learning the spectral-spatial features of HS image and the elevation features in LiDAR data. Subsequently, all features extracted from the two branches are fused at the feature level. Finally, they are combined with spectral-spatial and elevation information through decision-level fusion to achieve accurate classification. The experimental results on three real RS datasets demonstrate that our method exhibits better effectiveness and accuracy compared to existing technologies.

Text-to-speech (TTS), singing voice synthesis (SVS), and opera singing synthesis (OSS) aim to convert linguistic or musical inputs into intelligible and expressive speech. Although these tasks share a similar generative pipeline, they differ substantially in acoustic characteristics such as pitch range, prosody, and expressiveness. These distinctions pose significant challenges for unified modeling, particularly in learning representations that are both generalizable across tasks and sufficiently expressive for each task’s specific requirements. Existing approaches typically employ separate models for each task, resulting in increased computational overhead and limited cross-task knowledge transfer. To address these challenges, we propose UniVoice, a unified voice synthesis framework designed to jointly model TTS, SVS, and OSS. UniVoice adopts discrete acoustic units as a unified representation across tasks, simplifying the learning objective. To accommodate task-specific variability, we introduce the task-specific adapter with Top-1 routing that enables the acoustic model to specialize through lightweight, learnable adapters, effectively mitigating cross-task interference while preserving parameter efficiency. Furthermore, we incorporate a task-aware postnet that refines unit sequences based on task identity and musical context, enhancing both expressiveness and output quality. The final waveform is synthesized using a unit-based vocoder built on BigVGAN, conditioned on discrete units and pitch information. Experiments on multiple datasets spanning TTS, SVS, and OSS demonstrate that UniVoice significantly outperforms existing single-task and multi-task baselines. Specifically, UniVoice achieves improvements of 0.26, 0.12, and 0.22 in average Mean Opinion Scores (MOS) over the multi-task baseline for TTS, SVS, and OSS, respectively. Moreover, UniVoice achieves these gains with a compact model size of only 49.4M parameters and maintains a fast inference speed, indicating its effectiveness in balancing synthesis quality and computational efficiency.