Superior Generalization Capability Research Articles

Foundation models in robotics: Applications, challenges, and the future

We survey applications of pretrained foundation models in robotics. Traditional deep learning models in robotics are trained on small datasets tailored for specific tasks, which limits their adaptability across diverse applications. In contrast, foundation models pretrained on internet-scale data appear to have superior generalization capabilities, and in some instances display an emergent ability to find zero-shot solutions to problems that are not present in the training data. Foundation models may hold the potential to enhance various components of the robot autonomy stack, from perception to decision-making and control. For example, large language models can generate code or provide common sense reasoning, while vision-language models enable open-vocabulary visual recognition. However, significant open research challenges remain, particularly around the scarcity of robot-relevant training data, safety guarantees and uncertainty quantification, and real-time execution. In this survey, we study recent papers that have used or built foundation models to solve robotics problems. We explore how foundation models contribute to improving robot capabilities in the domains of perception, decision-making, and control. We discuss the challenges hindering the adoption of foundation models in robot autonomy and provide opportunities and potential pathways for future advancements. The GitHub project corresponding to this paper can be found here: https://github.com/robotics-survey/Awesome-Robotics-Foundation-Models .

Classification of cervical vertebral maturation stages with machine learning models: leveraging datasets with high inter- and intra-observer agreement

ObjectivesThis study aimed to assess the accuracy of machine learning (ML) models with feature selection technique in classifying cervical vertebral maturation stages (CVMS). Consensus-based datasets were used for models training and evaluation for their model generalization capabilities on unseen datasets.MethodsThree clinicians independently rated CVMS on 1380 lateral cephalograms, resulting in the creation of five datasets: two consensus-based datasets (Complete Agreement and Majority Voting), and three datasets based on a single rater’s evaluations. Additionally, landmarks annotation of the second to fourth cervical vertebrae and patients’ information underwent a feature selection process. These datasets were used to train various ML models and identify the top-performing model for each dataset. These models were subsequently tested on their generalization capabilities.ResultsFeatures that considered significant in the consensus-based datasets were consistent with a CVMS guideline. The Support Vector Machine model on the Complete Agreement dataset achieved the highest accuracy (77.4%), followed by the Multi-Layer Perceptron model on the Majority Voting dataset (69.6%). Models from individual ratings showed lower accuracies (60.4–67.9%). The consensus-based training models also exhibited lower coefficient of variation (CV), indicating superior generalization capability compared to models from single raters.ConclusionML models trained on consensus-based datasets for CVMS classification exhibited the highest accuracy, with significant features consistent with the original CVMS guidelines. These models also showed robust generalization capabilities, underscoring the importance of dataset quality.

An efficient corn leaf disease prediction using Adaptive Color Edge Segmentation with Resnext101 model

Corn leaf disease prediction is essential in ensuring agricultural productivity and food security. Although classical segmentation and classification techniques have been used to identify diseases, they often fail to deal with complex leaf structures. Thus, deep learning has attracted growing attention for automatically learning discriminative features from data. This paper proposes a new approach to predicting corn leaf disease using deep learning and classical segmentation techniques for improved accuracy and efficiency. The proposed work presents an adaptive colour edge segmentation method improved by a hill-climbing search for robust feature extraction according to the diversified nature of corn leaves. With the help of stochastic hill climbing, the segmentation process is optimized by adaptively changing parameters to get better segmentation accuracy. After that, features are extracted and provided as input into a ResNeXt-101 model for disease classification. The novel activation function in ResNeXt-101 combines the Adam optimizer with a Gaussian-based GELU to enhance nonlinearity to the smooth activation of GELU. The results obtained from experiments show that the proposed approach is practical. The ResNeXt-101 model has achieved overwhelming performance, with 99% for testing accuracy and a minimum loss of 0.01, proving its superior generalization capability and robustness. The proposed work was performed with no overfitting and free from local minima during the training process.

Analyzing the heterogenous effects of factors on high-range speeding likelihood of taxi speeders: Does explainable deep learning provides more insights than random parameter approach?

The random parameters Generalized Linear Model (GLM) is frequently used to model speeding characteristics and capture the heterogenous effects of factors. However, this statistical approach is seldom employed for prediction and generalization due to the challenge of transferring its predefined errors. Recently, the emergence of explainable AI techniques has illuminated a new path for analyzing factors associated with risky driving behaviors. Despite this, there remains a gap that comparing results from machine and deep learning (ML/DL) approaches with those from random parameters GLM. This study aims to apply the random parameter GLM and explainable deep learning to evaluate the heterogenous effects of factors on the taxis’ high-range speeding likelihood. Initially, a Beta GLM with random parameters (BGLM-RP) is developed to model the high-range speeding likelihood among taxi drivers. Additionally, XGBoost, a simple convolutional neural network (Simple-CNN), a deeper CNN (DCNN), and a deeper CNN with self-attention (DCNN-SA) are developed. The quantified explanations and illustrations of the factors’ heterogenous effects from ML/DL models are derived from pseudo coefficients by decomposing factors’ SHapley Additive exPlanations (SHAP) values. All the developed statistical, ML, and DL models are compared in terms of mean absolute errors and mean square errors on testing and full data. Results show that DCNN-SA excels in prediction on testing data, indicating its superior generalization capabilities, while BGLM-RP outperforms other models on full data. The DCNN-SA can reveal the heterogenous effects of factors for both in-sample and out-of-sample data, which is not possible for the random parameter GLM. However, BGLM-RP can reveal larger magnitudes of the factors’ heterogenous effects for in-sample data. The signs and significances are identical between the varying coefficients from BGLM-RP and the pseudo coefficients from the ML/DL models, demonstrating the validity and rationale of using the proposed explanation framework to quantify the factors’ effects in ML/DL models. The study also discusses the contributions of various factors to the high-range speeding likelihood of taxi drivers.

Continuous, Learned Imputation of Missing Values in Parkinson's Disease.

Parkinson's disease management requires accurate clinical scores but suffers from missing data. Leveraging self-supervised learning, we demonstrate superior generalization capabilities across populations compared to other well-established imputation techniques (MIWAE, MissForest, MICE). With the ability to employ the method already during the data collection and not afterward, the technology allows more robust data collection in clinical reality.

An analytic formulation of convolutional neural network learning for pattern recognition

Training convolutional neural networks (CNNs) using back-propagation (BP) is a time-consuming and resource-intensive process, primarily due to the need to iterate over the dataset multiple times. In contrast, analytic learning aims to train neural networks in a single epoch, offering a potential solution to these challenges. However, existing studies of analytic learning have been limited to multilayer perceptrons (MLPs). In this article, we propose an analytic formulation for convolutional neural network learning (ACnnL), which represents a significant advancement towards non-iterative learning paradigms for CNNs. Our formulation demonstrates that ACnnL extends the principles of MLP regularization constraints. From the implicit regularization and network interpretability viewpoints, we provide insights into why CNNs often exhibit superior generalization capabilities. The ACnnL is validated by conducting classification tasks on benchmark datasets such as MNIST, FashionMNIST, CIFAR10, CIFAR100 and Tiny-ImageNet. It is encouraging that the ACnnL trains CNNs in a significantly fast manner with reasonably close prediction accuracies to those using BP. In particular, a 5-layer vanilla CNN trained by ACnnL gave an accuracy of 0.9931, 0.9155, 0.7049 and 0.4628 for these datasets. The ACnnL achieves training speeds that are approximately 17 times faster than BP on GPU and 113 times faster than BP on CPU, while maintaining competitive prediction accuracies. Moreover, our experiments disclose a unique advantage of ACnnL under the small-sample scenario when training data are scarce or expensive. In a nutshell, an analytic method which deals well with small-sample size data has been put forward for the first time for fast CNN training with inherent network interpretability.

Soil Cadmium Prediction and Health Risk Assessment of an Oasis on the Eastern Edge of the Tarim Basin Based on Feature Optimization and Machine Learning

Soil heavy metal pollution poses a serious threat to food security, human health, and soil ecosystems. Based on 644 soil samples collected from a typical oasis located at the eastern margin of the Tarim Basin, a series of models, namely, multiple linear regression （LR）, neural network （BP）, random forest （RF）, support vector machine （SVM）, and radial basis function （RBF）, were built to predict the soil heavy metal content. The optimal prediction result was obtained and utilized to analyze the spatial distribution features of heavy metal contamination and relevant health risks. The outcomes demonstrated that： ① The average Cd content in the study area was 0.14 mg·kg-1, which was 1.17 times the soil background value of Xinjiang, making it the primary factor of soil heavy metal contamination in the area. Additionally, the carcinogenicity risk coefficients of Cd for both adults and children were less than 10-4, indicating that there were no significant long-term health risks for humans in the area. ② The estimation accuracies of the five inversion models were compared, and the validation set of the RF model had an R2 value of 0.763 7, which was the highest among the five models. Additionally, the RMSE, MAE, and MBE of the RF model were the smallest among the five models. Therefore, the predicted values of the RF model were most consistent with the measured values of the soil Cd content. The predicted map of soil Cd distribution derived from the RF model coincided best with the interpolation map. ③ The RF model outperformed the other four models in predicting health risks associated with the soil Cd element for both adults and children, resulting in better prediction results. Comparatively, the predicted values of the LR model in the validation set varied greatly, leading to unreliable results. It was demonstrated that the RF was the best model for predicting soil Cd content and evaluating health risks in the study area, considering its superior generalization capability and anti-overfitting ability.

MSRA-Net: multi-channel semantic-aware and residual attention mechanism network for unsupervised 3D image registration

Objective. Convolutional neural network (CNN) is developing rapidly in the field of medical image registration, and the proposed U-Net further improves the precision of registration. However, this method may discard certain important information in the process of encoding and decoding steps, consequently leading to a decline in accuracy. To solve this problem, a multi-channel semantic-aware and residual attention mechanism network (MSRA-Net) is proposed in this paper. Approach. Our proposed network achieves efficient information aggregation by cleverly extracting the features of different channels. Firstly, a context-aware module (CAM) is designed to extract valuable contextual information. And the depth-wise separable convolution is employed in the CAM to alleviate the computational burden. Then, a new multi-channel semantic-aware module (MCSAM) is designed for more comprehensive fusion of up-sampling features. Additionally, the residual attention module is introduced in the up-sampling process to extract more semantic information and minimize information loss. Main results. This study utilizes Dice score, average symmetric surface distance and negative Jacobian determinant evaluation metrics to evaluate the influence of registration. The experimental results demonstrate that our proposed MSRA-Net has the highest accuracy compared to several state-of-the-art methods. Moreover, our network has demonstrated the highest Dice score across multiple datasets, thereby indicating that the superior generalization capabilities of our model. Significance. The proposed MSRA-Net offers a novel approach to improve medical image registration accuracy, with implications for various clinical applications. Our implementation is available at https://github.com/shy922/MSRA-Net.

Research on the Maturity Detection Method of Korla Pears Based on Hyperspectral Technology

In this study, hyperspectral imaging technology with a wavelength range of 450 to 1000 nanometers was used to collect spectral data from 160 Korla pear samples at various maturity stages (immature, semimature, mature, and overripe). To ensure high-quality data, multiple preprocessing techniques such as multiplicative scatter correction (MSC), standard normal variate (SNV), and normalization were employed. Based on these preprocessed data, a custom convolutional neural network model (CNN-S) was constructed and trained to achieve precise classification and identification of the maturity stages of Korla pears. Additionally, a BP neural network model was used to determine the characteristic wavelengths for maturity assessment based on the sugar content feature wavelengths. The results demonstrated that the BP model, based on sugar content feature wavelengths, effectively discriminated the maturity stages of the pears. Specifically, the comprehensive recognition rates for the training, testing, and validation sets were 98.5%, 93.5%, and 90.5%, respectively. Furthermore, the combination of hyperspectral imaging technology and the custom CNN-S model significantly enhanced the detection performance of pear maturity. Compared to traditional CNN models, the CNN-S model improved the accuracy of the test set by nearly 10%. Moreover, the CNN-S model outperformed existing techniques based on partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) in capturing hyperspectral data features, showing superior generalization capability and detection efficiency. The superior performance of this method in practical applications further supports its potential in smart agriculture technology, providing a more efficient and accurate solution for agricultural product quality detection. Additionally, it plays a crucial role in the development of smart agricultural technology.

Bayesian optimization-based Model-Agnostic Meta-Learning: Application to predict maximum cyclic moment resistance of steel bolted T-stub connections

Accurately assessing the maximum moment resistance of steel bolted T-stub connections under cyclic loading is crucial for designing earthquake-resistant structures with such connections. Traditional methods based on design standards like ASCE 41-17 often lack precision. Recently, supervised machine learning techniques, particularly Artificial Neural Network (ANN), have been explored. However, conventional ANNs require substantial data for generalization, which is limited for steel bolted T-stub connections. To address these challenges, this study explores the feasibility of using the Model-Agnostic Meta-Learning (MAML) to predict the maximum cyclic moment resistance of steel bolted T-stub connections. MAML adapts task-specific model parameters rapidly and transfers knowledge across tasks to fine-tune global model parameters, potentially enhancing prediction accuracy with limited data. The MAML model is first optimized using Bayesian optimization with a Gaussian Process model to identify ideal hyperparameters. The optimized MAML model is then compared with two ANN models (one with optimized hyperparameters, another matching MAML's neural network architecture) and ASCE 41-17 method. Results demonstrate the optimized MAML model's superior generalization capabilities, offering a promising approach for steel bolted T-stub connections.

Benchmarking building energy consumption for space heating using an empirical Bayesian approach with urban-scale energy model

Calibration of an urban-scale building energy model is crucial for advancing building energy efficiency codes and policies to reduce carbon dioxide emissions whereas the accuracy of the model is challenged by performance gaps resulting from sparse data and uncertainties in modeling. An empirical Bayesian (EB) approach, incorporated a probability-based mixed-effects energy model and Monte Carlo Markov chain simulation, was developed to predict and benchmark building energy use intensity (EUI) for space heating, using the metering data of 2020 stochastically sampled district heating substations from a municipal utility database. The limitations, implementations and performances of different modeling strategy, including EB as well as Maximum Likelihood Estimation (MLE), Maximum A Posteriori (MAP) and Fully Bayes (FB) were compared. The results show that EB strategy achieved superior prediction performance than other strategies on the novel test dataset. The preparation, calibration, and validation efforts for implementing the EB approach were demonstrated. The proposed methodology is recommended as a robust approach for benchmarking urban-scale building EUI, distinguished from other machine learning approaches by requiring fewer training samples, better interpreting uncertainties in the data used, and exhibiting superior generalization capabilities. The proposed methodology was successfully demonstrated through the urban-scale benchmarking of space heating EUI in Beijing.

Split-and-Fit: Learning B-Reps via Structure-Aware Voronoi Partitioning

We introduce a novel method for acquiring boundary representations (B-Reps) of 3D CAD models which involves a two-step process: it first applies a spatial partitioning , referred to as the "split", followed by a "fit" operation to derive a single primitive within each partition. Specifically, our partitioning aims to produce the classical Voronoi diagram of the set of ground-truth (GT) B-Rep primitives. In contrast to prior B-Rep constructions which were bottom-up, either via direct primitive fitting or point clustering, our Split-and-Fit approach is top-down and structure-aware , since a Voronoi partition explicitly reveals both the number of and the connections between the primitives. We design a neural network to predict the Voronoi diagram from an input point cloud or distance field via a binary classification. We show that our network, coined NVD-Net for neural Voronoi diagrams, can effectively learn Voronoi partitions for CAD models from training data and exhibits superior generalization capabilities. Extensive experiments and evaluation demonstrate that the resulting B-Reps, consisting of parametric surfaces, curves, and vertices, are more plausible than those obtained by existing alternatives, with significant improvements in reconstruction quality. Code will be released on https://github.com/yilinliu77/NVDNet.

Detection Based on Semantics and a Detail Infusion Feature Pyramid Network and a Coordinate Adaptive Spatial Feature Fusion Mechanism Remote Sensing Small Object Detector

In response to the challenges of remote sensing imagery, such as unmanned aerial vehicle (UAV) aerial imagery, including differences in target dimensions, the dominance of small targets, and dense clutter and occlusion in complex environments, this paper optimizes the YOLOv8n model and proposes an innovative small-object-detection model called DDSC-YOLO. First, a DualC2f structure is introduced to improve the feature-extraction capabilities of the model. This structure uses dual-convolutions and group convolution techniques to effectively address the issues of cross-channel communication and preserving information in the original input feature mappings. Next, a new attention mechanism, DCNv3LKA, was developed. This mechanism uses adaptive and fine-grained information-extraction methods to simulate receptive fields similar to self-attention, allowing adaptation to a wide range of target size variations. To address the problem of false and missed detection of small targets in aerial photography, we designed a Semantics and Detail Infusion Feature Pyramid Network (SDI-FPN) and added a dedicated detection scale specifically for small targets, effectively mitigating the loss of contextual information in the model. In addition, the coordinate adaptive spatial feature fusion (CASFF) mechanism is used to optimize the original detection head, effectively overcoming multi-scale information conflicts while significantly improving small target localization accuracy and long-range dependency perception. Testing on the VisDrone2019 dataset shows that the DDSC-YOLO model improves the mAP0.5 by 9.3% over YOLOv8n, and its performance on the SSDD and RSOD datasets also confirms its superior generalization capabilities. These results confirm the effectiveness and significant progress of our novel approach to small target detection.

Predictive Study on the Occurrence of Wheat Blossom Midges Based on Gene Expression Programming with Support Vector Machines.

This study addresses the challenges in plant pest and disease prediction within the context of smart agriculture, highlighting the need for efficient data processing techniques. In response to the limitations of existing models, which are characterized by slow training speeds and a low prediction accuracy, we introduce an innovative prediction method that integrates gene expression programming (GEP) with support vector machines (SVM). Our approach, the gene expression programming-support vector machine (GEP-SVM) model, begins with encoding and fitness function determination, progressing through cycles of selection, crossover, mutation, and the application of a convergence criterion. This method uniquely employs individual gene values as parameters for SVM, optimizing them through a grid search technique to refine genetic parameters. We tested this model using historical data on wheat blossom midges in Shaanxi Province, spanning from 1933 to 2010, and compared its performance against traditional methods, such as GEP, SVM, naive Bayes, K-nearest neighbor, and BP neural networks. Our findings reveal that the GEP-SVM model achieves a leading back-generation accuracy rate of 90.83%, demonstrating superior generalization and fitting capabilities. These results not only enhance the computational efficiency of pest and disease prediction in agriculture but also provide a scientific foundation for future predictive endeavors, contributing significantly to the optimization of agricultural production strategies.

Moving toward Automaticity: A Robust Synthetic Occlusion Image Method for High-Throughput Mushroom Cap Phenotype Extraction

Mushroom cap is a key trait in the growth process and its phenotypic parameters are essential for automatic cultivation and smart breeding. However, the edible mushrooms are usually grown densely with mutual occlusion, which is difficult to obtain the phenotypic parameters non-destructively. Although deep learning methods achieve impressive performance with superior generalization capabilities, they require a large amount of ground truth label of the occluded target, which is a challenging task due to the substantial workload. To overcome this obstacle, a novel synthetic cap occlusion image method was proposed for rapidly generating edible mushroom occlusion datasets using raw images, in which the ground truth is obtained from the real world and the occlusion is randomly generated for simulating real scenes. Moreover, variants of amodal instance segmentation models with different backbone were trained and evaluated on our synthetic occlusion image datasets. Finally, an amodal mask-based size estimation method was presented to calculate the width and length of the cap. The experimental results showed that the amodal instance segmentation achieved an AP@[0.5:0.95] of 82%, 93% and 96% on Oudemansiella raphanipes, Agrocybe cylindraceas and Pholiota nameko synthetic cap datasets, respectively, with a size of 1024 × 1024 px, which indicates that our occlusion image synthesis method can effectively simulate the real cap occlusion situation. The size estimation method achieved an R2 of 0.95 and 0.98 between predictive amodal caps and manually labeled caps for the length and width of Agrocybe cylindraceas cap, respectively, which can be applied to obtain the phenotypic parameters of each cap effectively and accurately. These methods not only meet the demand for automatic monitoring of edible mushroom morphology in factories but also provide technical support for intelligent breeding.

Harnessing the potential of shared data in a secure, inclusive, and resilient manner via multi-key homomorphic encryption

In this manuscript, we develop a multi-party framework tailored for multiple data contributors seeking machine learning insights from combined data sources. Grounded in statistical learning principles, we introduce the Multi-Key Homomorphic Encryption Logistic Regression (MK-HELR) algorithm, designed to execute logistic regression on encrypted multi-party data. Given that models built on aggregated datasets often demonstrate superior generalization capabilities, our approach offers data contributors the collective strength of shared data while ensuring their original data remains private due to encryption. Apart from facilitating logistic regression on combined encrypted data from diverse sources, this algorithm creates a collaborative learning environment with dynamic membership. Notably, it can seamlessly incorporate new participants during the learning process, addressing the key limitation of prior methods that demanded a predetermined number of contributors to be set before the learning process begins. This flexibility is crucial in real-world scenarios, accommodating varying data contribution timelines and unanticipated fluctuations in participant numbers, due to additions and departures. Using the AI4I public predictive maintenance dataset, we demonstrate the MK-HELR algorithm, setting the stage for further research in secure, dynamic, and collaborative multi-party learning scenarios.

A deep learning approach for acute liver failure prediction with combined fully connected and convolutional neural networks.

Acute Liver Failure (ALF) is a critical medical condition with rapid development, often caused by viral infections, hepatotoxic drug abuse, or other severe liver diseases. Timely and accurate prediction of ALF occurrence is clinically crucial. However, predicting ALF poses challenges due to the diverse physiological differences among patients and the dynamic nature of the disease. This study introduces a deep learning approach that combines fully connected and convolutional neural networks for effective ALF prediction. The goal is to overcome limitations of traditional machine learning methods and enhance predictive model performance and generalization. The proposed model integrates a fully connected neural network for handling basic patient features and a convolutional neural network dedicated to capturing temporal patterns in patient data. The combination allows automatic learning of complex patterns and abstract features present in highly dynamic medical data associated with ALF. The model's effectiveness is demonstrated through comprehensive experiments and performance evaluations. It outperforms traditional machine learning methods, achieving 94.8% accuracy and superior generalization capabilities. The study highlights the potential of deep learning in ALF prediction, emphasizing the importance of considering individualized medical factors. Future research should focus on improving model robustness, addressing imbalanced data, and further exploring personalized features for enhanced predictive accuracy in real-world clinical scenarios.

Utilizing Data Science and AI for Customer Churn Prediction in Marketing

This study explores the application of data science and AI techniques in predicting customer churn within the telecommunications industry, a sector characterized by intense competition and high customer turnover rates. By analyzing historical customer data, including usage patterns and service preferences, the study aims to identify factors contributing to churn and propose targeted retention strategies to mitigate losses. Traditional classification algorithms and ensemble techniques are evaluated using the Telecom-Customer-Churn dataset, with emphasis on the underutilized Stacking ensemble method. The results demonstrate that ensemble learning algorithms, particularly the Stacking model, outperform single algorithms, with CatBoost exhibiting the highest accuracy at 0.8119, followed closely by RandomForest at 0.7902 and XGBoost at 0.7820. These findings underscore CatBoost's superior generalization capabilities, likely attributed to its adept handling of categorical features and missing values, and its ability to model complex data relationships. The study contributes to advancing understanding of ensemble models and offers valuable insights for predicting telecom customer churn, thereby aiding in the development of effective retention strategies and enhancing customer satisfaction and loyalty.

Learning context-aware local feature descriptors for 3D reconstruction

The generation of generalizable and discriminative descriptors plays a crucial role in image matching and 3D reconstruction. While numerous existing solutions are concentrated on encoding specific invariances, such as illumination or viewpoint invariance, they often face challenges in achieving robustness and generalization. These challenges arise from the frequent inadequacy of these solutions to effectively adapt to diverse and demanding environments due to their limited information capacity. In this paper, we introduce a novel approach aimed at maximizing the utilization of hidden feature informativeness to address these challenges. Specifically, we propose the Hierarchical Context-aware Aggregation Network (HCNet), which employs a hierarchical dense features constraint in a coarse-to-refinement description manner. In this approach, a coarse-level descriptor is used to present the overall information, while the refinement descriptor captures the detailed information of the image. Leveraging the strengths of both CNN and Transformer architectures, our hierarchical dense feature constraint encodes both local features and long-range information to efficiently generate dense feature descriptions. To boost descriptor informativeness and enhance matching accuracy, we introduce the Context-aware Attention Aggregation (CAA) model, which adaptively aggregates features from various scales through an efficient coarse-to-refinement manner. Additionally, we design a hierarchical triplet training strategy that considers both variant and invariant properties of hierarchical features, aiming to enhance descriptor informativeness while preserving their strong discriminative qualities. Our experiments, conducted on two popular feature-matching benchmarks, as well as a challenging long-term visual localization benchmark, demonstrate that our method significantly improves matching accuracy and outperforms state-of-the-art descriptors. Moreover, our approach exhibits superior generalization capabilities in various 3D reconstruction scenarios.

DFSA-DAN: dynamic fusion of statistical metric and adversarial learning for domain adaptation network based intelligent fault diagnosis

The advancement of deep transfer learning has motivated research into the realization of intelligent fault diagnosis schemes for rolling bearing. Nevertheless, existing research rarely provides further insight into the importance of statistical distance metric-based methods and adversarial learning-based methods in domain adaptation, and the commonly used feature extractors are more difficult to extract features suitable for domain transformation. In this paper, a dynamic fusion of statistical metric and adversarial learning for domain adaptation network is proposed to achieve a dynamic measure of the importance of different domain adaptation methods. This new model utilizes a local maximum mean discrepancy metric to adjust the conditional distribution and adversarial training to adjust the marginal distribution between domains. Meanwhile, to assess the importance of the two distributions, a dynamic adaptation factor is introduced for dynamic evaluation. In addition, to extract features that are more suitable for domain transformation, the model incorporates a dual depth convolutional path with an attention mechanism as a feature extractor, enabling multi-scale feature extraction. Experimental results demonstrate the model’s superior generalization capability and robustness, enabling effective cross-domain fault diagnosis in diverse scenarios.