Applied Intelligence最新文献

In occluded urban environments, the traditional object detection algorithm relies solely on RGB as input, making it challenging to discern the spatial relationship of occluded objects and consequently affecting the target detection accuracy. Previous studies primarily focused on fusing depth and RGB information at the feature level, resulting in the loss of detailed features from the original data, such as occlusion boundaries. This leads to blurred fusion features and degraded model detection performance. Therefore, this paper proposes a depth-guided RGB-D occluded target detection framework based on transformers (DGT) to effectively extract occlusion boundary information and guide the occlusion discrimination via data-level fusion of depth and RGB information. In particular, a multimodal data-level fusion model is proposed for a two-part task. One is to generate dense depth images with strengthened occlusion edge features by extracting the depth difference of object edges in the point cloud data. The other is to dilute the influence of useless information using RGB-D data-level fusion. A depth-guided occlusion layered detection network with transformers was designed to obtain the cross-module guided feature vector by exchanging the weights of the residual and interaction vectors. Extensive experiments showed that DGT achieves state-of-the-art performance in occluded environments.

In the process of monitoring spatiotemporal air quality data, data sample missingness is prevalent, thus rectifying missing values in spatiotemporal data holds paramount significance. In recent years, diffusion probability models have played a prominent role in image, video, and text generation, and have also begun to be applied in the field of spatiotemporal data imputation. However, such models face challenges in extracting fine-grained features for stable model operation and accurate modeling of data probability distributions. To address the aforementioned issues, we propose a Diffusion-based Generative adversarial framework for Multivariate air quality data Imputation, termed DGMI. Recognizing the similar temporal, sensor, and indicator change characteristics inherent in air quality data, our framework is designed to cater to the spatiotemporal characteristics of air quality data by incorporating a multi-cycle temporal feature extraction module and a sensor indicator feature extraction module, facilitating multidimensional refinement and integration of temporal, sensor, and indicator information. Moreover, the initial missing value is encoded with linear interpolation and sine-cosine functions. Following the generation of imputed values by the model, we introduce a discriminator module to discern the consistency between imputed values and observed values to provide feedback for optimizing the model from a data distribution perspective. DGMI outperforms most current data imputation methods under various missing ratios in two real air quality datasets by 4.1% (root mean square error) and 3.0% (mean absolute error), exhibiting efficacy in scenarios characterized by multidimensional spatiotemporal and high missing rates data.

This study focuses on image style transfer, aiming to generate images with the desired style while preserving the underlying content structure. Existing models face challenges in accurately representing both content and style features. To address this, an integrated method for image style transfer is proposed, utilizing a parallel CNN and Vision Transformer (CaVIT). It combines a Convolutional Neural Network (CNN) with a Vision Transformer (VIT) to achieve enhanced performance. Our method utilizes VGG-19 with residual blocks to encode style features for enhanced refinement. Additionally, the PA-Trans Encoder Layer is introduced, inspired by the Transformer Encoder Layer, to efficiently encode content features while preserving the complete content structure. The fused features are then decoded into stylized images using a CNN decoder. Qualitative and quantitative evaluations demonstrate that our proposed method outperforms existing models, delivering high-quality results.

Taking Shanghai as the research area, fully considering the randomness and timeliness of power battery recycling, a combined prediction model of Long Short-term Memory network—Support Vector Machine—Weibull (LSTM-SVM-Weibull) that integrates attention mechanism and hyperparameters optimized by improved student psychology based optimization (ISPBO) algorithm is proposed to predict the retired amount of power batteries, to further improve the prediction accuracy. In the first stage, the grey relational analysis (GRA) method is used to screen out the strong related influence factors of power battery installed amount. In the second stage, a two-stage predictive model of power battery installed amount based on LSTM network with attention mechanism (Attention-LSTM) and SVM optimized by ISPBO algorithm (ISPBO-SVM) is constructed. Firstly, the attention mechanism is fused in the hidden layer of the LSTM network to accurately predict the various indicators selected by GRA, reduce the impact of indicator value errors on the target value prediction, and highlight the contribution degree of the input sequence at different time prediction points; Then, the Lévy flight strategy is introduced and combined with the Metropolis criterion of the simulated annealing algorithm to accept inferior solutions, the ISPBO algorithm is designed to optimize the hyperparameters of SVM model, so as to predict the target value (power battery installed amount) on the basis of future indicator values by training the ISPBO-SVM target prediction layer. By collecting the historical data of power battery installed amount in Shanghai for simulation experiments, the comparison results show that the designed Attention-LSTM-ISPBO-SVM two-stage power battery installed amount prediction model is significantly better than other comparison models in terms of error and accuracy on different data sets, and it has high accuracy and generalization ability for the prediction of power battery retirement amount. In the third stage, based on the predictive model of power battery installed amount, the Weibull life distribution model is combined to predict the retired amount of power batteries in Shanghai. Through the goodness-of-fit test and evaluating the retirement amount prediction results of different models, it is proved that the predictive model for power battery retired amount based on Weibull life distribution has high stability and practical application value, which effectively reflects the overall change trend of the future power battery retirement amount in Shanghai, and can provide data reference for improving the recovery rate of waste batteries.

Formal context restoration is a recently developing topic in the field of formal concept analysis (FCA). Its goal is to restore a formal context from some known formal concepts. Each type of formal concept uses its unique perspective to restore formal contexts or concept lattices. This paper investigates the relationship among five types of basic concepts: the object concepts, the attribute concepts, the join-irreducible concepts, the meet-irreducible concepts, and the formal concepts in concept reducts. This paper first studies the elementary relationship among the basic sets of formal concepts and presents these relationship in a Venn diagram. Then the relationship among the basic concept sets are explored from a restoration perspective, specifically including the relationship among basic concepts from the perspectives of formal context restoration and concept lattice restoration. These relationship are then used to study the transformation between the basic concept sets, and are summarised in a formal context and its concept lattice. Finally, a practical case is given to illustrate the relationship among the basic concept sets explored in this paper, including the elementary relationship, and the relationship from the perspectives of formal context restoration and concept lattice restoration.

Fashion image retrieval emphasizes accurately perceiving the fine-grained features to meet users’ precise needs. However, the existing global image-based retrieval methods encounter challenges such as imprecise positioning of attributes, difficulty in distinguishing visually similar but semantically different attribute values, and struggles in the learning of attribute features within specific regions and viewpoints. This paper proposes a two-stage hybrid framework called IPAD (Iterative Positioning and Attribute Diverging) for attribute-aware fashion similarity learning. In the initial stage, we present an iterative positioning strategy to precisely identify local attribute regions through an iterative attention mechanism with adaptive suppression. IPAD leverages the strengths of Convolutional Neural Networks and Vision Transformers. Subsequently, we design an attribute diverging strategy to optimize attribute value aggregation via online clustering using a momentum encoder, thereby enhancing model stability and representation. During inference, we further present a feature reasoning mechanism to refine retrieval results through subgraph similarity matrix generation and re-ranking to enhance accuracy and robustness. Extensive evaluations on three public datasets demonstrate IPAD’s superior performance over state-of-the-art methods in retrieval accuracy, achieving an average improvement in MAP by +4.22%. The source code is available at https://github.com/h8e9r7/IPAD.

Detecting suicidal ideation on communication platforms such as social media is critical for suicide prevention, as these platforms are frequently used for emotional expression and can reflect significant behavior changes. Many machine learning and deep learning techniques have been employed to address this issue, utilizing embedding methods such as Count Vector, Term Frequency-Inverse Document Frequency, Bidirectional Encoder Representations from Transformers, Multilingual Universal Sentence Encoder etc generate high-dimensional vectors. Directly inputting word embeddings into models can introduce noise and outliers, which may negatively impact predictive accuracy. Therefore, feature selection to optimize the dimensionality of word embedding vectors has emerged as a promising direction for future research. This study proposes a feature selection method called Propose Best Feature Selection, which combines Grey Wolf Optimization, Recursive Feature Elimination, and Stepwise Feature Selection. It uses a Voting Classifier to identify and filter the most significant features, reducing dimensionality. These optimized features are then fed into a stacked ensemble hybrid model, with Bi-Directional Gated Recurrent Unit with Attention and Convolutional Neural Network, acting like base and Extreme Gradient Boostis working like the meta-classifier, achieving an accuracy of 98% in Reddit and 97% in Twitter(X) dataset, outperforming similar methods in the field. This work is focused on textual data, and future efforts may expand to include multimodal analysis, incorporating image-based emotional cues. Scalability challenges for large datasets and real-time applications remain a key limitation.

Path planning is essential for simulating crowd evacuation. However, existing path planning methods encounter challenges, including unbalanced exit utilization, ineffective obstacle avoidance, and low evacuation efficiency. To address these issues, this paper presents a path planning method based on Deep Reinforcement Learning (DRL) and a Repulsive Force Field (RFF) for crowd evacuation simulation. First, a dynamic exit scoring mechanism is proposed and integrated into the DRL training process to balance exit utilization during evacuation. Additionally, we address the sparse reward issue in DRL by extracting key points from actual evacuation trajectories as short-term goals. Finally, we enhance the movement strategy output by constructing an RFF to improve obstacle avoidance in complex environments. Experimental results demonstrate that the proposed method effectively avoids obstacles and efficiently completes evacuation tasks.

Medical image registration is a fundamental task in medical image analysis. Recently, competitive methods, such as deep learning-based registration and deformable registration, which have demonstrated promising results, have been proposed. However, meeting the demands for high precision in clinical applications is still a challenge. Here, we propose a cyclic optimization registration framework that deeply fuses diffeomorphic and deep learning methods through a single forward-two path structure. A neural network estimates the initial deformation field, which directly generates registered images to enhance feature extraction capabilities. Additionally, dynamic diffeomorphism is introduced for the initial deformation field to generate the final deformation field, ensuring the invertibility of the transformation. We incorporate the Dense Spatial Correspondence Prompt module for cyclically learning the final deformation field, enabling the network to estimate smoother and more accurate spatial transformations. Experiments conducted on a publicly available 3D dataset demonstrate exceptional registration accuracy with a DSC of 0.621 and an SSIM of 0.913, while preserving desirable diffeomorphic properties with almost zero non-positive Jacobians.