Applied Intelligence最新文献

Shadow detection for transparent objects is a challenging task. The difficulty arises from the fact that transparent objects and shadow regions are prone to occlusion, and the boundaries of transparent objects become more blurred due to optical effects, ultimately leading to incomplete shadow detection results. In this paper, a novel semisupervised shadow detection algorithm based on self-correction is proposed to address these problems. We construct a shadow detection module based on a hybrid attention mechanism CBAM and integrate the short-term memory capability of LSTM networks, assisting the model in accurately localizing shadow areas based on prior experience. To address the issue of easily overlooked shadow areas, we aim to minimize the difference between the predicted shadow mask and the real shadow mask as our optimization objective. We train the shadow self-correction module using binary cross-entropy loss to enhance the model’s ability to detect shadow areas that are prone to be overlooked. Furthermore, a pretrained boundary detector is utilized to obtain the boundary information between the predicted and real shadow masks. The shadow detection model is then optimized under the constraint of boundary consistency, enabling the model to more accurately identify the boundaries of shadow regions and enhancing the shadow detection performance. The experimental results indicate that, compared to existing shadow detection algorithms, the proposed algorithm performs well in terms of both transparent and nontransparent object shadow detection.

The study focuses on applying machine learning methodologies to football player data for predicting player market values in the dynamic football market. Player datasets are rich, encompassing performance metrics, physiological attributes, and contextual variables. Machine learning models, including both traditional and advanced methods, effectively extract insights from complex data to estimate player market values. Addressing challenges like overfitting and computational complexity involves applying regularization, feature engineering, and interpretability tools to manage high-dimensional data and improve predictive accuracy. In this study sensitivity of selected models (Support Vector Regression (SVR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGB), and Categorical Boosting (CAT)) models to extracted data from FIFA 19 and Real-world Statistical Datasets evaluated by Shapley Additive Explanations (SHAP) and the 20 most relevant features selected in the ranking of SHAP for each regression model. Then, models optimized with two meta-heuristic algorithms demonstrated their performance in predicting the market values of players. Dempster-Shafer Theory (DST) was utilized to develop an ensemble of models to overcome overfitting problems, and Fourier amplitude sensitivity testing (FAST) gave insight for future data extractions. The analysis of market values for players revealed significant model performance variations. XGSC hybrid model demonstrated exceptional precision with a minimal error of 1.7 million dollars (10% of average measured value), followed by RSCX_SC with misestimations of 2 million dollars (13.3% of average measured value). Extracted results suggested that models, especially ensemble form, offer reliable accuracy for club managers and stakeholders, aiding in strategic player selection based on previous performance. This approach proves particularly beneficial for optimizing player salaries, especially when considering a prominent team with market values above average.

While numerous batch-mode active learning (BMAL) methods have been developed for nominal classification, the absence of a BMAL method tailored for ordinal classification is conspicuous. This paper focuses on proposing an effective BMAL method for ordinal classification and argues that a BMAL method should guarantee that the selected instances in each iteration are highly informative, diverse from labeled instances, and diverse from each other. We first introduce an expected model output change criterion based on the kernel extreme learning machine-based ordinal classification model and demonstrate that the criterion is a composite containing both informativeness assessment and diversity assessment. Selecting instances with high scores of this criterion can ensure that the selected are highly informative and diverse from labeled instances. To ensure that the selected instances are diverse from each other, we propose a leadership tree-based batch instance selection approach, drawing inspiration from density peak clustering algorithm. Thus, our BMAL method can select a batch of peak-scoring points from different high-scoring regions in each iteration. The effectiveness of the proposed method is empirically examined through comparisons with several state-of-the-art BMAL methods.

Parasitic infections pose a significant health risk in many regions worldwide, requiring rapid and reliable diagnostic methods to identify and treat affected individuals. Recent advancements in deep learning have significantly improved the accuracy and efficiency of microscopic image analysis workflows, enabling its application in various domains such as medical diagnostics and microbiology. This work presents DT4PEIS, a novel two-stage architecture for the instance segmentation of parasite eggs in microscopic images. The first stage is a DEtection TRansformer (DETR) based architecture, which predicts the bounding boxes and class labels of the detected eggs. Then, the predicted bounding boxes are used as prompts to guide the segmentation process in the second stage, which is based on the Segment Anything Model (SAM) architecture. We evaluate the performance of the proposed method on the Chula-ParasiteEgg-11 dataset. Our results show that the proposed method outperforms the other architectures in terms of segmentation mean Average Precision (mAP), providing a more detailed and accurate representation of the detected eggs.

High-speed and highly dynamic hypersonic vehicles demand exceptional safety and reliability during flight. Accurate detection and localization of faults in actuators and reaction control systems are pivotal for controlling and predicting operational states. However, this process encounters challenges such as multiple fault modes, limited data availability, and suboptimal diagnostic accuracy. Our focus is on common fault types in reaction control systems and actuators. We have designed a residual module and an attention module to construct an interpretable fault diagnosis model that extracts deep features from fault residual sequences and state parameter sequences. This model allows for the simultaneous and precise identification of fault type, location, and occurrence time. Furthermore, we visualize the diagnosis process through the use of attention weights and class activation mapping, thereby enhancing the interpretability of the fault diagnosis and bolstering the reliability of the results. Our findings reveal that both the residual module and attention module enhance diagnostic accuracy. In the diagnosis network, shallow attention primarily facilitates feature fusion, whereas deep attention primarily serves to filter features and improve detection capabilities. Without increasing computational complexity, the interpretable fault diagnosis model achieved an accuracy of 96.65%, and the fault time localization error was reduced by 86.15%. The proposed method simplifies model training and elevates fault detection accuracy, offering a reliable approach for isolating and identifying actuator faults.

Underwater object detection has numerous applications in protecting, exploring, and exploiting aquatic environments. However, underwater environments pose a unique set of challenges for object detection including variable turbidity, colour casts, and light conditions. These phenomena represent a domain shift and need to be accounted for during design and evaluation of underwater object detection models. Although methods for underwater object detection have been extensively studied, most proposed approaches do not address challenges of domain shift inherent to aquatic environments. In this work we propose a data-centric framework for combating domain shift in underwater object detection with image enhancement. We show that there is a significant gap in accuracy of popular object detectors when tested for their ability to generalize to new aquatic domains. We used our framework to compare 14 image processing and enhancement methods in their efficacy to improve underwater domain generalization using three diverse real-world aquatic datasets and two widely used object detection algorithms. Using an independent test set, our approach superseded the mean average precision performance of existing model-centric approaches by 1.7–8.0 percentage points. In summary, the proposed framework demonstrated a significant contribution of image enhancement to underwater domain generalization.

In numerous real-world applications, data tends to be ordered and partially labelled, predominantly due to the constraints of labeling costs. The current methodologies for managing such data are inadequate, especially when confronted with the challenge of high-dimensional datasets, which often require reprocessing from the start, resulting in significant inefficiencies. To tackle this, we introduce an incremental semi-supervised feature selection algorithm that is grounded in neighborhood discernibility, and incorporates pseudolabel granular balls and matrix updating techniques. This novel approach evaluates the significance of features for both labelled and unlabelled data independently, using the power of neighborhood distinguishability to identify the most optimal subset of features. In a bid to enhance computational efficiency, especially with large datasets, we adopt a pseudolabel granular balls technique, which effectively segments the dataset into more manageable samples prior to feature selection. For high-dimensional data, we employ matrices to store neighborhood information, with distance functions and matrix structures that are tailored for both low and high-dimensional contexts. Furthermore, we present an innovative matrix updating method designed to accommodate fluctuations in the number of features. Our experimental results conducted across 12 datasets-including 4 with over 2000 features-demonstrate that our algorithm not only outperforms existing methods in handling large samples and high-dimensional datasets but also achieves an average time reduction of over six fold compared to similar semi-supervised algorithms. Moreover, we observe an average improvement in accuracy of 1.4%, 0.6%, and 0.2% per dataset for SVM, KNN, and Random Forest classifiers, respectively, when compared to the best-performing algorithm among the compared algorithms.

Economic dispatching of power system includes real-time dispatching and day-ahead dispatching. In this process, accurate real-time and day-ahead load forecasting is crucial. However, integrating real-time forecasting and day-ahead forecasting into one system, and ensuring that both have good performance, is a challenging problem. To solve the above problem, we propose a load forecasting system based on deep learning and weighted ensemble. The system is composed of the high precision prediction module and the intelligent weighted ensemble module. In the high precision prediction module, we use variational mode decomposition (VMD) to decompose the data into multiple components of different frequencies, and build a selection pool that includes statistical models and deep learning to select the best prediction model for each component through customed metrics. In the intelligent weighted ensemble module, we improve the Grey Wolf optimization algorithm with tent chaos mapping and flight strategy. The improved Grey Wolf optimization algorithm (ILGWO) is used to determine the weight of each component, then the weight is multiplied by the component prediction result, and the final prediction result is obtained by adding. To verify the superiority of the proposed forecasting system, we conducted experiments using four sets of load data from New South Wales, Australia. Through six groups of experiments and three groups of discussion, the accuracy, stability and applicability of the load forecasting system are verified. Compared with the traditional method, the prediction accuracy (MAPE) of the proposed load forecasting system is improved by about 55%. In addition, we further validated the generality of the system with four sets of load data from Queensland, Australia. The results show that the proposed load forecasting system is significantly superior to other models and provides more reliable load forecasting for power system management and scheduling.

Maintaining high-precision localization and ensuring map consistency are crucial objectives for mobile robots. However, loop closure detection remains a challenging aspect of their operation because of viewpoint and appearance changes. To address this issue, this paper proposes WP-VLAD, a novel hierarchical loop closure detection method that tightly couples global features and weighted local patch-level features (WPs). WP-VLAD employs MobileNetV3 as the backbone network for feature extraction, and integrates a trainable vector of local aggregated descriptors (VLAD) for compact global and local feature representation. A hierarchical navigable small world method is used to retrieve loop candidate frames based on the global features, whereas a multiscale feature fusion weighted map prediction module assigns weights to the local patches during mutual nearest neighbour matching. The proposed weight allocation strategy emphasizes salient regions, reducing interference from dynamic objects. The experimental results on benchmark datasets demonstrate that WP-VLAD significantly improves matching performance while maintaining efficient computation, exhibiting strong generalizability and robustness across various complex environments.