Ships are a fundamental element of water transport traffic scenarios and the primary focus of waterway traffic monitoring. Shipping transportation, as a predominant mode of transportation, has witnessed rapid development in recent years. The automated classification of inland river ships serves as the foundation for the digitization and intelligent management of inland waterway transportation. It is crucial for facilitating the high-quality development of the shipping industry. The predominant approach for inland ship classification relies on visual sensors and synthetic aperture radar, which are limited in providing detailed 3D geometric information and are affected by varying weather and lighting conditions. In this paper, we propose a LiDAR-based ship classification method for inland waterways to address this issue. This method involves background filtering and target detection on the original point cloud, generating a dataset of point clouds of inland ships, and using PointNet to learn and classify ship point cloud features. Moreover, for the first time, we propose a point cloud classification framework for multi-channel feature fusion. The proposed framework fuses LiDAR ring ID, intensity, and geometric features into a unified point cloud representation. Based on the fused point cloud data, an improved model with a point-wise attention mechanism is employed for feature extraction and classification. Our method achieves an accuracy of 97.33%, surpassing the geometric information-only method by 2.94%. This result effectively demonstrates the method’s efficacy in extracting features and classifying LiDAR point cloud ships.

Modality imbalance is a significant challenge for multi-modal interaction at various depths in multispectral pedestrian detection under varying illumination environments. To overcome the limitations of current cross attention in addressing the modality imbalance, we propose the Cross-Modal Dual-Stream Feature Interaction Transformer (CDFIT). CDFIT capitalizes on the Transformer’s ability to learn long-range dependencies, extracting global intra-modal and inter-modal correlations during the feature interaction phase. Crucially, in order to effectively eliminate the interference of the self-attention within one modality to the alternative one, we propose horizontal and vertical correlation decoupling modes to divide and reassemble the attention maps in CDFIT. This facilitates more purified inter-modal attention while preserving relevant intra-modal self-attention, reducing the information interference. Meanwhile, in CDFIT, we expand Transformer into dual-stream pathways to align and assemble the information from RGB and thermal modalities across depths separately, thereby greatly enhancing the performance of multispectral object detection. Comprehensive experiments and ablation studies on benchmark datasets demonstrate that CDFIT achieves superior performance compared with state-of-the-art methods.

In the field of 3D vision, 3D point cloud completion is a critical task in many practical applications. This study proposes a point cloud completion network designed to mitigate the impact of uneven point cloud distribution on completion tasks. By combining explicit neighborhood aggregation methods with implicit association techniques, the network achieves balanced feature representation across regions with varying distributions. Our approach comprises a Distribution-Geometry Feature Extractor (DGFE), a Seed Generator (SG), and a Point Generator (PG). DGFE leverages the proposed Dynamic Differential Distribution-Aware Module (D3AM) to process and enhance point cloud features layer by layer. SG generates seed point clouds and their corresponding features, while PG utilizes Cross-Resolution Upsampling Blocks (CRUB) to progressively generate denser point clouds by integrating point clouds and features across different resolutions. Experimental results demonstrate that our method achieves state-of-the-art performance on the PCN, ShapeNet-55/34, and KITTI benchmark datasets.

Urban rail transit (URT) is an essential part of urban public transportation. Accurate ridership prediction is increasingly important for the safe operation and efficient management of URT. However, existing studies regard the URT stations with different intersecting subway lines as a whole, which ignores the internal spatial connections within the stations. In fact, URT stations are embodiments of spatial coupling between subway lines. Additionally, the intrinsic patterns of ridership are also neglected. To further improve the prediction accuracy, this study proposes a deep learning model based on graph convolutional network (GCN) and bidirectional long short-term memory network (Bi-LSTM) with a non-parallel structure (D-BLGCN). At the beginning, this study decouples the URT stations according to the intersecting subway lines. On the basis of spatial decoupling, different patterns of ridership are diverged into tributaries. Then, a non-parallel structure in the proposed model is designed to capture the intrinsic spatio-temporal correlations of ridership. To the best of our knowledge, this is the first time that the integration of internal spatial connections and ridership diverging is employed for URT ridership prediction. Extensive experiments are conducted on Beijing URT ridership data with different time granularities. The results demonstrate that the proposed model achieves better prediction performance compared with baselines.

To interprete the importance of clinical features and genotypes for warfarin daily dose prediction, we developed a post-hoc interpretable framework based on an ensemble predictive model. This framework includes permutation importance for global interpretation and local interpretable model-agnostic explanation (LIME) and shapley additive explanations (SHAP) for local explanation. The permutation importance globally ranks the importance of features on the whole data set. This can guide us to build a predictive model with less variables and the complexity of final predictive model can be reduced. LIME and SHAP together explain how the predictive model give the predicted dosage for specific samples. This help clinicians prescribe accurate doses to patients using more effective clinical variables. Results showed that both the permutation importance and SHAP demonstrated that VKORC1, age, serum creatinine (SCr), left atrium (LA) size, CYP2C9 and weight were the most important features on the whole data set. In specific samples, both SHAP and LIME discovered that in Chinese patients, wild-type VKORC1-AA, mutant-type CYP2C9*3, age over 60, abnormal LA size, SCr within the normal range, and using amiodarone definitely required dosage reduction, whereas mutant-type VKORC1-AG/GG, small age, SCr out of normal range, normal LA size, diabetes and heavy weight required dosage enhancementt.

Traffic sign detection and recognition (TSDR) has attracted extensive studies recently due to its broad application prospect in Intelligent Transport Systems. TSDR is still challenging due to the small size of traffic signs in the image. Besides, the traffic signs in the real world exhibit a long-tailed distribution (i.e., data for most categories are scarce while for others are abundant.), which will lead to a significant performance drop of the detection framework. In this paper, we propose a novel traffic sign detection framework to address these challenging problems. In order to detect small traffic signs, we propose an effective adaptive and attentive spatial feature fusion module which learns the spatial attention map to fuse different feature maps at each scale while emphasizing or suppressing the features at different regions. This module can significantly alleviate the inconsistency among features and enhance feature representations of small objects. Furthermore, to address the long-tailed data problem, a hierarchical group softmax head which constructs a label tree to divide categories into different groups is proposed, in this way, categories in each group have relatively similar frequencies, then the softmax is applied in each relatively balanced group to calculate the probability of each category. Extensive experiments conducted on the TT100K and GTSDB datasets demonstrate that the proposed method achieves notable improvement in both the small traffic signs and long-tailed detection problems in TSDR.

To improve the accuracy of arrhythmia diagnosis and reduce the recheck time, we design a cascaded step-temporal attention network called ArrhythmiaNet to classify 15 categories of arrhythmias on electrocardiogram (ECG) signals. In ArrhythmiaNet, the first level contains a convolution layer with step-attention, which recognizes abnormal heartbeat and provides morphological feature expression. The second level is composed of a gated recurrent unit (GRU) with temporal attention, which mines the temporal correlation of long-term rhythm for abnormal rhythm judgement. To share the feature expression, ArrhythmiaNet was trained by end-to-end multitask learning. In the experiment, ArrhythmiaNet and the comparison algorithms were tested on a dataset (819 training samples and 264 test samples) from MIT-BIH arrhythmia database. The results showed that the accuracy of ArrhythmiaNet was 20.3% higher than that of support vector machine (SVM), Naive Bayesian, gradient boost decision tree (GBDT) and random forest (RF), and 8.2% higher than that of long-term memory network (LSTM) and recurrent neural network (RNN). Compared with 1-dimension convolutional neural network (1D-CNN), ArrhythmiaNet obtained similar overall accuracy, higher recall and precision. Compared to the genetic ensemble of SVM classifiers and evolutionary neural system, ArrhythmiaNet has much lower complexity than them with competitive accuracy. Besides, ArrhythmiaNet has higher interpretability in arrhythmias diagnosis.

To solve the data imbalance problem and improve the predictive accuracy on warfarin daily dosage, we develop an evolutionary synthetic minority oversampling technique (ESMOTE), which is based on an evolutionary strategy (ES). ESMOTE oversamples the minority, whose genotypes are *1/*3 and *3/*3 for CYP2C9, AG and GG for VKORC1 or who took amiodarone and drank regularly. Ensemble learning method-Random Forest (RF) is used to build ensemble predictive model. RF produces a group of trees by training them on minority and the masses as well as different combinations of features. And, they make uses of correlation of tress to improve the generalization of predictive model. In the experiment, five machine learning methods are as comparators to ESMOTE-RF. These methods are tested on the inner dataset of The First Affiliated Hospital of Soochow University and an external dataset of International Warfarin Pharmacogenetics Consortium (IWPC). Results showed that ESMOTE-RF present the highest accuracy on the prediction of warfarin dose in terms of R-squared (R2) and mean squared error (mse). In terms of the percentage of patients whose predicted dose of warfarin is within 20% of the actual stable therapeutic dose (20%-p), ESMOTE-RF can achieve satisfied prediction.

An evolutionary ensemble modeling (EEM) method is developed to improve the accuracy of warfarin dose prediction. In EEM, genetic programming (GP) evolves diverse base models, and the genetic algorithm optimizes the parameters of the GP. The EEM model is assembled by using the prepared base models through a technique called “bagging.” In the experiment, a dataset of 289 Chinese patients, which was provided by the First Affiliated Hospital of Soochow University, is used for training, validation, and testing. The EEM model with selected feature groups is benchmarked with four machine-learning methods and three conventional regression models. Results show that the EEM model with the M2+G group, namely age, height, weight, gender, CYP2C9, VKORC1, and amiodarone, presents the largest coefficients of determination (R2), the highest percentage of the predicted dose within 20% of the actual dose (20%-p), the smallest mean absolute error, mean squared error, and root-mean-squared error on the test set, and the least decrease in R2from the training set to the test set. In conclusion, the EEM method with M2+G delivers superior performance and can, therefore, be a suitable prediction model of warfarin dose for clinical applications.

The prediction of daily stable warfarin dosage for a specific patient is difficult. To improve the predictive accuracy and to build a highly accurate predictive model, we developed an ensemble learning method, called evolutionary fuzzy c-mean (EFCM) clustering algorithm with support vector regression (SVR). A dataset of 517 Han Chinese patients was collected from the data of The First Affiliated Hospital of Soochow University and dataset of International Warfarin Pharmacogenetics Consortium for training and testing. In EFCM+SVR, we adopted SVR to build a generalized base model (SVR model). To achieve an accurate prediction on patients with large dosage, we proposed an EFCM clustering algorithm that can be used to cluster the training set and designed a clustering model on clusters and centroids. The SVR and clustering models were integrated into an ensemble model by stepwise functions. In the experiment, three artificial neural networks, SVR, two ensemble models, and three regression models were used as comparators to the EFCM+SVR model, which obtained the smallest mean absolute error (0.67 mg/d) in warfarin dose prediction and the largest R-squared (43.9%). The model achieved satisfactory prediction in terms of the percentage of patients whose predicted dose of warfarin was within 15% and 20% of the actual stable therapeutic dose (15%-p of 36% and 20%-p of 46.6%).

Most published results show that power reduction of the finite-state machines (FSMs) is achieved by decomposition. In order to achieve a low power FSM implementation, a Genetic Fuzzy c-mean clustering-based decomposition method, called GFCM-D, is proposed for FSM partition in this study. GFCM-D used Fuzzy c-mean clustering (FCM) to partition a set of states of FSM into a collection of c fuzzy clusters, then a FSM is decomposed into several sub machines. For achieving low power, the objective function of GFCM-D is to minimize the cross state transition probability between sub machines and increase the inner state transition probability within the submachine. Genetic algorithm (GA) is used as a shell, which applies selection, crossover and mutation for generating better centers and more appropriate clusters. We have tested our approach, GFCM-D, extensively on fifteen benchmarks, comparing it with previous FSM synthesis methods from various aspects. The experimental results show that GFCM-D has achieved a significant cost reduction of both dynamic power and leakage power dissipation over the previous publications.

In this study, we propose a module-level Evolvable Hardware (EHW) approach to design synchronous sequential circuits and minimize the circuit complexity (the number of logic gates and wires used). Firstly, we use Genetic Algorithm (GA) to implement state simplification and obtain near-optimal state assignment. Then, in the pre-evolution stage, EHW evolves a set of high performing circuits and uses data mining method to find frequently evolved blocks from these circuits. The frequently evolved block would be re-used as function or terminal for evolving better circuits in the re-evolution stage. EHW has a faster convergence so that the circuit with small complexity could be evolved. Auto starting ability of circuits would also be test by the fitness function of EHW. Finally, sequence detectors, modulon counters, and ISCAS'89 circuit are used as the proof for our evolutionary design approach. Simulation results are given, and our evolutionary algorithm is shown to be better than other methods in terms of convergence time, success rate, and maximum fitness across generations.