Feature Transfer Learning Research Articles

Enhanced YOLOv5 algorithm for helmet wearing detection via combining bi-directional feature pyramid, attention mechanism and transfer learning

Enhanced YOLOv5 algorithm for helmet wearing detection via combining bi-directional feature pyramid, attention mechanism and transfer learning

Rainstorm-Induced Emergency Recognition from Citizens’ Communications Based on Spatial Feature Extraction and Transfer Learning

The rapid recognition of rainstorm-induced emergencies (e.g., building and road inundation, facility damage, and trapped citizens) is vital to timely disaster response. One big challenge that limits the performance of emergency recognition is the data imbalance between different emergency domains. The present study aims to develop an effective cross-domain transfer learning framework for rainstorm-induced emergency recognition based on the text reports provided by citizens. The critical component of the framework is the use of joint distribution adaption (JDA) analysis embedded in a discriminative feature mapping procedure, which transfers rich knowledge learned from large-scale datasets to the learning task from small emergence data. Considering the feature incompleteness that is caused by short text length, a basic probability assignment function is constructed and applied to extract important spatial features for rainstorm emergency recognition, with an improved marginal Fisher analysis being adopted to optimize cross-domain text feature representation. The proposed scheme is validated using the empirical data of ten emergency classes from Wuhan City, China. Our experimental results show that the proposed method could significantly address data imbalance and thus help achieve high recognition performance through domain knowledge complementation. Meanwhile, the use of various spatial features is proved to be effective in tackling missing features. This scheme can be further developed into smart systems for rainstorm disaster response with reasonable performance and imbalanced sample sizes.

Model-Based Transfer Reinforcement Learning Based on Graphical Model Representations.

Reinforcement learning (RL) plays an essential role in the field of artificial intelligence but suffers from data inefficiency and model-shift issues. One possible solution to deal with such issues is to exploit transfer learning. However, interpretability problems and negative transfer may occur without explainable models. In this article, we define Relation Transfer as explainable and transferable learning based on graphical model representations, inferring the skeleton and relations among variables in a causal view and generalizing to the target domain. The proposed algorithm consists of the following three steps. First, we leverage a suitable casual discovery method to identify the causal graph based on the augmented source domain data. After that, we make inferences on the target model based on the prior causal knowledge. Finally, offline RL training on the target model is utilized as prior knowledge to improve the policy training in the target domain. The proposed method can answer the question of what to transfer and realize zero-shot transfer across related domains in a principled way. To demonstrate the robustness of the proposed framework, we conduct experiments on four classical control problems as well as one simulation to the real-world application. Experimental results on both continuous and discrete cases demonstrate the efficacy of the proposed method.

Predicting Cardiovascular Rehabilitation of Patients with Coronary Artery Disease Using Transfer Feature Learning.

Cardiovascular diseases represent the leading cause of death worldwide. Thus, cardiovascular rehabilitation programs are crucial to mitigate the deaths caused by this condition each year, mainly in patients with coronary artery disease. COVID-19 was not only a challenge in this area but also an opportunity to open remote or hybrid versions of these programs, potentially reducing the number of patients who leave rehabilitation programs due to geographical/time barriers. This paper presents a method for building a cardiovascular rehabilitation prediction model using retrospective and prospective data with different features using stacked machine learning, transfer feature learning, and the joint distribution adaptation tool to address this problem. We illustrate the method over a Chilean rehabilitation center, where the prediction performance results obtained for 10-fold cross-validation achieved error levels with an NMSE of 0.03±0.013 and an R2 of 63±19%, where the best-achieved performance was an error level with a normalized mean squared error of 0.008 and an R2 up to 92%. The results are encouraging for remote cardiovascular rehabilitation programs because these models could support the prioritization of remote patients needing more help to succeed in the current rehabilitation phase.

Multi-objective dynamic distribution adaptation with instance reweighting for transfer feature learning

In knowledge adaptation, the source and target knowledge are transferred into the same mapping space by simultaneously reducing the difference between the marginal and conditional distributions; however, it is difficult to precisely balance the two distributions at each transformation. To address this problem, a novel multi-objective dynamic distribution adaptation (MODDA) with instance reweighting is proposed to reduce discrepancies between the two distributions. In addition, a customised non-dominated sorting genetic algorithm-II (NSGA2) optimisation method is presented for searching the optimal cumulative weight path, and four genetic operator combinations are compared to determine which one is optimal for MODDA. Moreover, kernel mean matching is proposed for the first time for dynamic compensation based on an individual’s relevance in instance reweighting. The experimental results confirm that MODDA outperforms other state-of-the-art algorithms in terms of the classification accuracy for 16 well-known cross-domain tasks.

A novel feature and sample joint transfer learning method with feature selection in semi-supervised scenarios for identifying the sequence of some species with less known genetic data

A novel feature and sample joint transfer learning method with feature selection in semi-supervised scenarios for identifying the sequence of some species with less known genetic data

Probability-based Graph Embedding Cross-domain and Class Discriminative Feature Learning for Domain Adaptation.

Feature-based domain adaptation methods project samples from different domains into the same feature space and try to align the distribution of two domains to learn an effective transferable model. The vital problem is how to find a proper way to reduce the domain shift and improve the discriminability of features. To address the above issues, we propose a unified Probability-based Graph embedding Cross-domain and class Discriminative feature learning framework for unsupervised domain adaptation (PGCD). Specifically, we propose novel graph embedding structures to be the class discriminative transfer feature learning item and cross-domain alignment item, which can make the same-category samples compact in each domain, and fully align the local and global geometric structure across domains. Besides, two theoretical analyses are given to prove the interpretability of the proposed graph structures, which can further describe the relationships between samples to samples in single-domain and cross-domain transfer feature learning scenarios. Moreover, we adopt novel weight strategies via probability information to generate robust centroids in each proposed item to enhance the accuracy of transfer feature learning and reduce the error accumulation. Compared with the advanced approaches by comprehensive experiments, the promising performance on the benchmark datasets verify the effectiveness of the proposed model.

Condition Monitoring of Wind Turbine Generators Based on SCADA Data and Feature Transfer Learning

In order to build an effective condition monitoring (CM) model for the target wind turbines (WTs) with few operational data, an approach based on the feature transfer learning and a modified generative adversarial network is proposed. First, a large amount of labelled data from WTs are analyzed to construct a CM model with the aid of an autoencoder. This forms the knowledge of CM for WTs in the source domain. Second, a generative adversarial network is trained to build a mapping relationship between the features of different WTs. Third, the health status of the target WT is determined by analyzing the data collected from it online based on the proposed approach. Two case studies are conducted to verify that the proposed method can transfer the CM knowledge from source WT to target WT and achieve good performance in the CM of target WT.

Feature fusion based artificial neural network model for disease detection of bean leaves

<abstract> <p>Plant diseases reduce yield and quality in agricultural production by 20–40%. Leaf diseases cause 42% of agricultural production losses. Image processing techniques based on artificial neural networks are used for the non-destructive detection of leaf diseases on the plant. Since leaf diseases have a complex structure, it is necessary to increase the accuracy and generalizability of the developed machine learning models. In this study, an artificial neural network model for bean leaf disease detection was developed by fusing descriptive vectors obtained from bean leaves with HOG (Histogram Oriented Gradient) feature extraction and transfer learning feature extraction methods. The model using feature fusion has higher accuracy than only HOG feature extraction and only transfer learning feature extraction models. Also, the feature fusion model converged to the solution faster. Feature fusion model had 98.33, 98.40 and 99.24% accuracy in training, validation, and test datasets, respectively. The study shows that the proposed method can effectively capture interclass distinguishing features faster and more accurately.</p> </abstract>

High Resolution Remote Sensing Image Classification Based on Deep Transfer Learning and Multi Feature Network

High Resolution Remote Sensing Image Classification Based on Deep Transfer Learning and Multi Feature Network

Baby Cry Recognition by BCRNet Using Transfer Learning and Deep Feature Fusion

Baby Cry Recognition by BCRNet Using Transfer Learning and Deep Feature Fusion

Deep Multiple Instance Learning Approach for Classification in Clinical Decision Support Systems

To get around the drawbacks of conventional classification algorithms that required manual feature extraction and the high computational cost of neural networks, this paper introduces a deep convolutional neural network with multiple instance learning approaches, namely dynamic max pooling and sparse representation. For the categorization of tuberculosis lung illness, this model combines deep convolutional neural networks and multiple instance learning. The design was composed of four phases: pre-processing, instance production, feature extraction, and classification. To perform feature extraction, a model based on a customized version of the VGG16 architecture was trained from scratch. Multiple instance learning techniques such as Diverse Density (DD) and the Maximum pattern bag formulation of the Support Vector Machine were used to evaluate how well the proposed classification algorithm performed in comparison (SVM).The numerical findings demonstrated that the new method offered a higher level of accuracy than the methods that had been used in the past. When evaluating the efficacy of the current method, accuracy, specificity, sensitivity, and error rate were all taken into consideration. The accuracy of the max-pooling based framework and the sparse representation framework was found to be greater than that of the other multiple instance strategies, coming in at 91.51% and 89.84%, respectively, when compared to that of the other methods. The improved accuracy of the present system that makes use of deep neural networks is mostly attributable to the contributions made by features such as transfer learning and automatic feature extraction.

A Novel Hybrid Transfer Learning Approach for Small-Sample High-Voltage Circuit Breaker Fault Diagnosis On-site

Although the data-driven fault diagnosis method has achieved perfect diagnosis of high-voltage circuit breakers (HVCBs) mechanical fault under the massive data built in the laboratory, it is still a challenge to train a high-precision and robust diagnosis model under the condition of small samples on-site at this stage. To solve the above issues, this paper proposes a novel hybrid transfer learning to realize small-sample HVCB fault diagnosis on-site. To fully learn domain discriminative features and domain matching, this paper simultaneously introduces domain adaptation transfer learning and domain adversarial training into small-sample HVCB diagnosis on-site. At the same time, the two kinds of feature transfer learning are combined through ensemble learning to get the final diagnosis result. To extract discriminative features that characterize HVCB faults, this paper constructs a one-dimensional attention residual convolutional neural network, which can ensure that the network pays attention to key features while fully extracting temporal fine-grained information. The experimental results show that the hybrid transfer learning proposed in this paper achieves 94.69% accuracy of small-sample HVCB fault diagnosis on-site, which is significantly higher than other methods. It has laid a solid foundation for small-sample HVCB fault diagnosis on-site.

Adaptive Fusion Feature Transfer Learning Method For NILM

Non-invasive load monitoring (NILM) aims to extract the power consumption of individual appliances from a smart meter that measures the total power consumption of all appliances. At present, deep learning methods have achieved leading results. However, the need for a large number of training data and the poor generalization ability of models limit the further development of NILM. To break through these limitations, this paper proposes an adaptive fusion feature transfer learning method. Firstly, to provide rich features for feature transfer, multiple feature extraction branches are used to extract temporal and spatial features from different perspectives. These features are fused for adaptive adjustment. Secondly, the attention mechanism is used to adapt the features extracted from the model of the source task so that they are more conducive to the model of the target task. Finally, abundant simulation experiments are performed for appliance transfer and house transfer, respectively. It is verified that the proposed method can achieve better results than the existing method by using only a small amount of training data and retraining a part of the new model.

Ultrasound images-based deep learning radiomics nomogram for preoperative prediction of RET rearrangement in papillary thyroid carcinoma.

To create an ultrasound -based deep learning radiomics nomogram (DLRN) for preoperatively predicting the presence of RET rearrangement among patients with papillary thyroid carcinoma (PTC). We retrospectively enrolled 650 patients with PTC. Patients were divided into the RET/PTC rearrangement group (n = 103) and the non-RET/PTC rearrangement group (n = 547). Radiomics features were extracted based on hand-crafted features from the ultrasound images, and deep learning networks were used to extract deep transfer learning features. The least absolute shrinkage and selection operator regression was applied to select the features of nonzero coefficients from radiomics and deep transfer learning features; then, we established the deep learning radiomics signature. DLRN was constructed using a logistic regression algorithm by combining clinical and deep learning radiomics signatures. The prediction performance was evaluated using the receiver operating characteristic curve, calibration curve, and decision curve analysis. Comparing the effectiveness of the models by linking the area under the receiver operating characteristic curve of each model, we found that the area under the curve of DLRN could reach 0.9545 (95% confidence interval: 0.9133-0.9558) in the test cohort and 0.9396 (95% confidence interval: 0.9185-0.9607) in the training cohort, indicating that the model has an excellent performance in predicting RET rearrangement in PTC. The decision curve analysis demonstrated that the combined model was clinically useful. The novel ultrasonic-based DLRN has an important clinical value for predicting RET rearrangement in PTC. It can provide physicians with a preoperative non-invasive primary screening method for RET rearrangement diagnosis, thus facilitating targeted patients with purposeful molecular sequencing to avoid unnecessary medical investment and improve treatment outcomes.

Center-based Transfer Feature Learning With Classifier Adaptation for surface defect recognition

Surface defect recognition using Deep Learning based computer vision techniques is an important task in industrial manufacturing. However, surface images have different distributions due to different environments in industrial manufacturing, where the distribution difference between images will degrade the accuracy of Deep Learning based computer vision techniques. In addition, existing Transfer Feature Learning (TFL) methods reduced the distribution difference only by the location parameters of distributions, which ignored the effect of scale parameters in representing the distribution. To overcome these problems, we propose Center-based Transfer Feature Learning with Classifier Adaptation (CTFLCA) for surface defect recognition. First, to eliminate the distribution difference at the feature layer from the location parameters and scale parameters of distributions, we utilize centers as bases to propose the Center-based Transfer Feature Learning method (CTFL) by minimizing Center-based Distribution Difference of Location Parameters (CDDLP) and Center-based Distribution Difference of Scale Parameters (CDDSP). Second, to reduce the distribution difference at the classifier layer from the location parameters and scale parameters, we establish the Center-based Classifier Adaptation method (CCA) using a similar idea of CTFL, where the optimization objective of CCA is formulated by minimizing the classification errors, CDDLP in classification results, and CDDSP in classification results. Next, under the guidance of the class-wise sample selection, we establish CTFLCA for integrating CTFL with CCA. Finally, sufficient results on four datasets (NEU-C, PR-C, Office-Caltech, and ImageCLEF-DA) illustrate the effectiveness of CTFLCA, where the average classification accuracies of CTFLCA are 99.6%, 98.0%, 95.1%, and 91.1%.

A Wind Speed Combination Forecasting Method Based on Multifaceted Feature Fusion and Transfer Learning for Centralized Control Center

• A novel combined model based on transfer learning is proposed for wind speed forecasting in different data features and scenarios. • MVMD is used to decompose the multivariate series with the intention of considering the internal coupling relationship between wind speed and meteorological series. • TDCN-BiLSTM model is established for alleviating the interference between data, extracting multi-scales characteristics. • The IMOGOA is developed to improve the optimization ability of the algorithm. With the establishment of remote centralized control centers for wind farms, the combined model of multiple deep neural networks used in most wind speed prediction methods can no longer meet the requirements of centralized control centers for efficient and low-cost wind speed prediction in different locations. Thus, a model based on multi-layer feature fusion and transfer learning is proposed in this paper for multi-location wind speed forecasting. Under the idea of "offline training and online prediction", the wind speed characteristics of wind farms at typical locations are learned in the offline stage by fusing multifaceted features through a two-channel dilated convolutional network and a bi-directional long short-term memory network, transferring the well-trained models to any target wind farm to achieve online prediction, and integrating the prediction results of each offline model with an improved multi-objective grasshopper optimization algorithm to further enhance the prediction accuracy. Finally, three comparisons reveal that the proposed model outperforms other baseline models in adaptability and accuracy.

K-Means Clustering Approach to UAS Classification via Graphical Signal Representation of Radio Frequency Signals for Air Traffic Early Warning

Small Unmanned Aerial Systems (UASs) provide significant benefits to economies across the world on a daily basis but increased usage brings a number of security challenges. For example, identifying small UASs operating with malicious intent in a restricted airspace. Supervised learning techniques applied to radio frequency (RF) signals have been considered for the classification of UAS type with high accuracy but due to labelled data assume the UAS signal is already known. Unsupervised learning algorithms such as K-means clustering provide a potential for identifying small UAS signals which have not been seen before. The use of transfer learning and CNN feature extraction (FE) with spectrogram graphical signal representations have been successfully used in a supervised manner for medical diagnosis and audio classification. This research is the first application of transfer learning and CNN FE as a pre-cursor to an unsupervised learning algorithm. This paper shows that clustering graphical representations of the signal and utilising CNN FE with transfer learning produces the highest v-measure score 0.814 but at a cost of 6s in time. Small UASs can travel at speeds of 45 mph so timely detection is essential in many use cases. A decrease in 0.2 v-measure score using PSD graphical image representations of the RF signal and PCA initialisation allows the clustering time to complete in under 0.3s even in environments with active interference in the same band. This timely result could provide effective early warning with the cueing of a secondary sensor or supervised algorithm with higher classification accuracy.

Social distance and face mask detector system exploiting transfer learning

<p class="IJASEITAbtract"><span lang="EN-US">As time advances, the use of deep learning-based object detection algorithms has also evolved leading to developments of new human-computer interactions, facilitating an exploration of various domains. Considering the automated process of detection, systems suitable for detecting violations are developed. One such applications is the social distancing and face mask detectors to control air-borne diseases. The objective of this research is to deploy transfer learning on object detection models for spotting violations in face masks and physical distance rules in real-time. The common drawbacks of existing models are low accuracy and inability to detect in real-time. The MobileNetV2 object detection model and YOLOv3 model with Euclidean distance measure have been used for detection of face mask and physical distancing. A proactive transfer learning approach is used to perform the functionality of face mask classification on the patterns obtained from the social distance detector model. On implementing the application on various surveillance footage, it was observed that the system could classify masked and unmasked faces and if social distancing was maintained or not with accuracies 99% and 94% respectively. The models exhibited high accuracy on testing and the system can be infused with the existing internet protocol (IP) cameras or surveillance systems for real-time surveillance of face masks and physical distancing rules effectively.</span></p>

Multi-source satellite imagery and point of interest data for poverty mapping in East Java, Indonesia: Machine learning and deep learning approaches

This study proposes a novel approach to provide a more granular poverty map in terms of coverage (up to a grid level with the spatial resolution of 1.5 km) with less cost and time to update to support better poverty monitoring. Two poverty estimation model development scenarios using machine learning and deep learning were evaluated in this study. First, the model is constructed using multisource satellite imagery and geospatial point of interest location of economic infrastructures using zonal statistics feature extraction, which is the proposed approach in this study. The specific multisource satellite imagery to be used as poverty indicators are night-time light intensity (NTL) as an approach to economic activity, normalized difference vegetation index (NDVI) as an approach to detect rural areas based on vegetation, built-up index (BUI) as an approach to detect urban areas based on building distribution, normalized difference water index (NDWI) to detect land cover, land surface temperature (LST) as an approach to detecting urban areas based on surface temperature, carbon monoxide (CO), nitrogen dioxide (NO2), and sulfur dioxide (SO2) as an approach to detecting economic activities based on pollution, as well as point of interest (POI) density, and POI distance which show the accessibility of an area. Second, the model is constructed from daytime multiband and nighttime light intensity satellite imagery using the deep learning architecture of Resnet-34 transfer learning feature extraction. In each scenario, we compared the performance of different machine learning (support vector regression/SVR, decision tree regression/DTR, and random forest regression/RFR) algorithms along with deep learning (multiple layer perceptron/MLP and one dimensional-convolutional neural network/CNN-1D) algorithms. Based on the results of the model development evaluation, the CNN-1D model was chosen as the best model in the first scenario, and the Resnet-34 + MLP model was chosen as the best model in the second scenario. The two best models from the first and second scenarios are then used to predict poverty at the grid level with a spatial resolution of 1.5 km. Based on official poverty data, the estimated poverty mapping built using the CNN-1D model in the first scenario was chosen as the best map with the value of root mean squared error (RMSE) 1.95 and adjusted R2 0.84 at the district level. The results of visual identification revealed that high estimates of poverty are typically found in sparsely inhabited areas surrounded by unoccupied land, which is usually an agricultural area. Otherwise, low poverty estimates, on the other hand, are more likely to be found in highly populated regions with convenient access. Our finding is aligned with the official report of Statistics Indonesia, which claims that poverty in rural regions is higher than in urban areas.