Applied Intelligence最新文献

Traditional streamflow forecasting models based on machine learning often struggle to determine optimal parameters and achieve accurate predictions when applied to small and medium-sized rivers with limited data availability. To address this issue, this study proposes an Attention-based Bidirectional Long Short-Term Memory (A_BiLSTM) and, on this basis, develops a streamflow forecasting model based on model fusion, referred to as SFM_MF. The SFM_MF model employs Bayesian Linear Regression (BLR) to learn from small-sample data in the target basin, while leveraging the A_BiLSTM to model data transferred from hydrologically similar source basins. The final streamflow forecast is generated by integrating the outputs of BLR and A_BiLSTM through a weighted averaging method. Experimental results from the Jiulong River Basin and the Qinhuai River Basin demonstrate that A_BiLSTM outperforms baseline models, and that the SFM_MF model significantly surpasses its component models in terms of both predictive accuracy and generalization capability. Compared to the individual models, SFM_MF achieves reductions in root mean square error (RMSE) of 3.98% and 19.33%, respectively. These findings indicate that the SFM_MF model delivers superior forecasting performance for small and medium-sized basins with limited data resources.

The vehicle routing problem with time windows (VRPTW) has gained much attention recently due to its wide application in operations research and logistics. VRPTW has been proven to be an NP-hard problem whose optimal solution is computationally costly. Scholars have proposed many methods, such as exact algorithms, heuristics, and metaheuristics, to find near-optimal solutions for the VRPTW. Exact algorithms are limited to small-scale problems, while heuristic algorithms and metaheuristics often converge to locally optimal solutions, despite their applicability to larger-scale problems. This paper proposes a novel membrane-inspired evolutionary algorithm framework (MEAF) consisting of isolated evolutionary rules, communication output rules, communication input rules, fusion-exchange information operation, and membrane dissolution rules. By leveraging the advantages of multiple metaheuristics algorithms and avoiding the pitfalls of local optima, MEAF offers a promising solution to address complex problems. The effectiveness of the proposed MEAF is verified by applying three classical metaheuristics, namely Genetic Algorithm (GA), Ant Colony System (ACS), and Particle Swarm Algorithm (PSO), to solve the VRPTW problem. The experiments are run on 56 instances of Solomon with 100 client benchmarks. The evaluation of the experimental results combined with the mean and standard deviation values show that the algorithm performs better in 54 out of 56 instances, demonstrating the effectiveness and stability of the proposed algorithm.

Industrial product surface defect detection plays a crucial role in guaranteeing the quality of industrial products. Several primary challenges remain in achieving efficient and accurate automatic detection: (1) Industrial product surface defects demonstrate irregular morphological features. (2) Significant size disparities between different defects result in the risk of information loss during feature fusion in Feature Pyramid Networks (FPN) typically employed in the neck component. (3) Visual similarity among different defect categories means that slight deviations in predicted box locations can result in misclassification errors. This research initially employs a DCN-C3k2 module design, incorporating DCNv3 to improve the model’s sensitivity to varied morphological information of objects. Furthermore, we design an Adaptive Focusing Diffusion Network (AFDN) that aggregates multi-scale features and implements adaptive channel selection and fusion, then propagates critical features through upsampling and downsampling processes to improve defect detection precision. Lastly, we introduce a Task Dynamic Interactive Detection Head (TDIDH). The TDIDH constructs detection head architecture according to the specific properties and variations of different detection tasks, with the objective of optimizing detection performance through enhanced inter-task dynamic interaction. Experiments are performed on public datasets GC10-DET and NEU-DET. The proposed method achieves AP, AP(_{50}), and AP(_{75}) scores of 41.7%, 77.1%, and 40.8% respectively on NEU-DET, demonstrating improvements of 3.7%, 4.5%, and 4.1% compared to YOLOv11. Results on GC10-DET show scores of 44.3%, 81.5%, and 39.8%, with improvements of 3.3%, 0.9%, and 3.3% respectively compared to YOLOv11. Additionally, we achieve a 50% reduction in both parameter count and computational load through model pruning while preserving detection accuracy equivalent to the original AFFNet model, facilitating deployment on edge devices. The experimental results validate the effectiveness of our proposed method. The code is released at: https://github.com/yifansdut/affnet.

Malignancy prediction of lung nodules on computed tomography (CT) is a task of significant clinical importance. The malignancy of lung nodules on CT is not only reflected in some local details but also closely related to several global attributes. However, previous studies often ignore this vital fact, resulting in a bottleneck of performance improvement. In this paper, a collaborative deep learning framework based on adaptive feature fusion (CDLF-AFF) is proposed to improve the malignancy prediction performance. In the CDLF-AFF method, a multi-view input strategy is designed to drive the pre-trained models to capture the abundant spatial information of nodule CT images. Furthermore, a dual-branch architecture is developed to simultaneously learn the local structure features and long-range dependency relations. To improve the feature fusion efficiency, a feature aggregation module is constructed to adaptively fuse the feature maps produced by two different styles of learning branches. The CDLF-AFF method is evaluated on three different datasets. On the benchmark dataset LIDC-IDRI, it achieved an AUC of 97.25% (95% CI: 96.97%-97.54%), representing improvements of 1.47% and 0.86% over the corresponding unimodal models, ResNet-Model and ViT-Model, respectively. On the clinical dataset CQUCH-LND, it achieved an AUC of 94.09% (95% CI: 93.34%-94.84%), showing improvements of 3.41% and 1.39% over the corresponding ResNet-Model and ViT-Model, respectively. On the competition dataset LUNGx, it achieved an AUC of 79.13% (95% CI: 78.07%-80.19%), surpassing the best-performing algorithm listed on the competition leaderboard by 11.13%. These results demonstrate that the CDLF-AFF can effectively predict the malignancy of lung nodules.

This research presents a novel energy-efficient task sequencing method for manufacturing operations involving multiple processing points, such as precision drilling and contact welding. The problem is formulated as a multi-dimensional weighted variant of the Traveling Salesman Problem (TSP), and solved using a Multi-gate Mixture of Experts (MMOE) neural architecture. Unlike previous approaches that require separate models for each TSP size, our method employs a single neural network to handle TSPs of all sizes, significantly improving scalability and reducing training overhead. With an uncertainty-based loss weighting strategy, the model effectively balances multiple learning objectives. Experiments show that MMOE-9 achieves performance comparable to state-of-the-art methods with only one-third of the parameters of NAR4TSP, and its training time is similar to that of a single TSP100 model. Further, we extend the model to cover 91 TSP sizes (from 10 to 100) within the same unified framework, demonstrating strong generalization across scales.

The person search task aims to address both person detection and person re-identification(re-id) simultaneously, integrating these two tasks into a unified objective. Person search is commonly used in surveillance and security fields. Currently, person search tasks in surveillance scenarios face many severe challenges, such as scale variations and occlusion issues caused by cameras. Existing approaches often overlook the discrepancies between multi-scale features and typically perform direct feature fusion. Most methods addressing occlusion rely on feature completion techniques, without fully utilizing the inherent fine-grained information from the original images. This paper proposes a Multi-level Supervised and Fine-grained Feature Enhancement for Person Search (MFPS) to mitigate these issues. MFPS employs cascaded encoders and decoders to extract person detection features from the backbone network. To generate re-id features robust to scale variations, MFPS introduces a Multi-Level Supervision method (MLS), which aggregates features of different scales and levels, enriching the semantic information of person features. Furthermore, to address the issue of missing re-id features caused by occlusion, this paper proposes a deformable fine-grained attention module. This module extracts fine-grained re-id features with accurate semantic information through sampling point offset operations. Finally, fine-grained features and multi-scale features are fused, and the re-id features extracted through multi-level supervised fine-grained feature extraction significantly improve recognition accuracy for person search tasks in surveillance scenarios. The experimental results show that MFPS improves the mAP metrics by 0.8 and the top-1 metrics by 1.8 compared to the state-of-the-art method on the PRW dataset, proving its superiority in complex environments. The source code is available at https://github.com/FengHua0208/MFPS.

Social recommendation utilizes social connections as auxiliary information to deeply mine user preferences, thereby improving the recommendation performance. Existing methods often employ graph neural networks to encode social graphs. However, average aggregation may lead to node distortion, and diffusion models may reconstruct embeddings of multi-faceted interests and attributes that are misaligned with task-relevant directions. To address these limitations, this paper introduces a Conditional Guided Diffusion architecture in Latent Space for social recommendation (CGDLS). Specifically, CGDLS first leverages singular value decomposition (SVD) to encode the social connection graph and the co-interacted items graph among users into the low-dimensional latent space. To alleviate noise distortion, CGDLS captures its key features by leveraging SVD to encode the social connection graph. During the reverse process, CGDLS incorporates the embedding of co-interacted items among users as conditional guidance. It guides the reverse process to reconstruct a highly task-relevant social connection embedding. Extensive experiments conducted on three social datasets demonstrate that CGDLS and its components outperform various state-of-the-art methods.

Among men globally, prostate cancer ranks as one of the most frequently occurring types of cancer. Timely identification and accurate treatment rely on high-quality medical image registration, especially the precise registration of multimodal images such as diffusion-weighted imaging (DWI) and T2-weighted MRI. However, due to the low resolution, high noise and blurred boundaries of DWI, the registration with T2-weighted images is extremely challenging. The significant differences between the two imaging modalities make the multimodal registration task complex, which directly affects the accuracy of diagnosis and treatment. To address these problems, this paper proposes a dense convolutional and gated feature extraction network (DG-Morph), which leverages the strengths of Convolutional Neural Networks (CNN) and Transformer models to enhance the efficiency of feature representation and extraction. The gated residual fusion module (GRFM) dynamically fuses features, the Multi-scale Transformer module (MST) improves multi-scale feature extraction, and the dense 3D-convolutional(Dense Conv3D) block optimizes feature reuse. Experimental results show that DG-Morph significantly outperforms traditional methods regarding the Dice Similarity Coefficient(DSC) and deformation field smoothness, showing excellent performance in complex multimodal medical image registration. These results show that DG-Morph has significant advantages in accuracy and robustness in multimodal registration, especially for complex multimodal medical image registration tasks.

Financial markets are highly non-linear and noisy, which complicates stock price forecasting. Traditional machine learning models often fail to appropriately generalize noisy data. Additionally, incorporating redundant attributes increases training complexity and overfitting risk. In this study, a hybrid model is proposed that combines noise reduction and feature selection techniques to improve forecasting accuracy and efficiency. Specifically, we propose a fusion noise identification and attribute reduction method based on independent component analysis (ICA), signal-to-noise ratio (SNR) theory, and neighborhood rough set (NRS) theory to address the aforementioned problems. First, technical indicators are established and used as alternative inputs for ICA. The aim is to determine the optimal ICA input dimension and noise sequence by maximizing the SNR while minimizing the reconstructed similarity of the series. Second, we calculate the attribute significance using the neighborhood rough set and obtain an attribute subset that satisfies the threshold filter condition. Finally, an attention mechanism-based (AM) long short-term memory (LSTM) neural network is used to build a forecasting model for the noise-reduced close price and reduced set of attributes. The research findings indicate that for time series denoising based on ICA, optimal results are achieved with an appropriate input dimension. Excessive input sequences introduce noise or redundant information. Attribute reduction based on NRS effectively reduces both temporal and spatial computational complexity without sacrificing generalization accuracy.

Siamese-based tracking approaches have demonstrated notable efficacy in visual object tracking recently. Nevertheless, existing Siamese networks suffer from core limitations in their feature extraction and interaction mechanisms: namely, a lack of adaptive modeling capability for multi-scale context and low sensitivity to target instance variations. These shortcomings cause trackers to be prone to drifting when encountering challenges such as occlusion, scale variations, and fast motion. To address this problem, we develop an original dynamic multi-order context fusion siamese network for object tracking, constructing a three-level progressive functional coupling architecture. By introducing the Multi-Order Feature Gated Fusion module, the Adaptive Fine-grained Channel Cross-correlation module, and the Spatial-Channel Coordinated Attention module, we effectively enhance discriminative representation learning, cross-correlation matching accuracy, and target activation. We conducted extensive experiments on four benchmark datasets: OTB100, UAV123, GOT-10K, and LaSOT, to confirm our tracker’s superior performance. Our code is available at: https://github.com/JSJ515-Group/SiamDMCF.