Transfer Learning Method Research Articles

A New Hybrid Model for Artificial Intelligence Assisted Tire Defect Detection: CTLDF+EnC

This paper focuses on an artificial intelligence based worn tire detection system proposed to detect cracks in the tires of vehicle drivers. Although drivers are generally aware of the importance of tire tread depth and air pressure, they are not aware of the risks associated with tire oxidation. However, tire oxidation and cracks can cause significant problems affecting driving safety. In this paper, we propose a new hybrid architecture for tire crack detection, CTLDF+EnC (Cascaded Transfer Learning Deep Features + Ensemble Classifiers), which uses deep features from pre-trained transfer learning methods in combination with ensemble learning methods. The proposed hybrid model utilizes features from nine transfer learning methods and classifiers including Stacking, Soft and Hard voting ensemble learning methods. Unlike X-Ray image-based applications for industrial use, the model proposed in this study can work with images obtained from any digital imaging device. Among the models proposed in the study, the highest test accuracy value was obtained as 76.92% with the CTLDF+EnC (Stacking) hybrid model. With CTLDF+EnC (Soft) and CTLDF+EnC (Solid) models, 74.15% and 72.92% accuracy values were obtained respectively. The results of the study show that the proposed hybrid models are effective in detecting tire problems. In addition, a low-cost and feasible structure is presented.

Inception ResNet v2 for Early Detection of Breast Cancer in Ultrasound Images

Breast cancer is one of the leading causes of death in women. Early detection through breast ultrasound images is important and can be improved using machine learning models, which are more accurate and faster than manual methods. Previous research has shown that the use of the CNN (Convolutional Neural Network) algorithm in breast cancer detection still does not achieve high accuracy. This study aims to improve the accuracy of breast cancer detection using the Inception ResNet v2 transfer learning method and data augmentation. The data is divided into training, validation and testing data consisting of 3 classes, namely Benign, Malignant and Normal. The augmentation process includes rotation, zoom, and rescale. The model trained using CNN and Inception ResNet v2 showed good performance by producing the highest accuracy of 89.72% in the training data evaluation data and getting 90% accuracy in the prediction test stage with data testing. This study shows that the combination of data augmentation and the Inception ResNet v2 architecture can improve the accuracy of breast cancer detection in CNN models.

Automatic Classification of Melanoma Skin Cancer Images with Vision Transform Model and Transfer Learning

Melanoma is one of the most aggressive and lethal forms of skin cancer. Therefore, early diagnosis and correct diagnosis are very important for the health of the patient. Diagnostic procedures require human expertise, increasing the possibility of error. With developing technology, advances in deep learning models have become hope for the automatic detection of Melanoma skin cancer with computer systems. The Vision Transformer (ViT) model was developed by Google and has achieved very successful results in the field of classification. In this study, the transfer learning method was applied with the ViT model using the melanoma skin cancer dataset taken from the Kaggle library and the performance of the model was evaluated. Before starting training, pre-processing was applied to the data set. The dataset consists of 9600 training and 1000 test images. Training and experimental testing of the model was carried out with Python language on the Colab platform. As a result of the experimental studies conducted on the test data set, it was seen that the model reached an accuracy rate of 93.5% and was competitive with existing models

Unsupervised Transfer Learning Method via Cycle-Flow Adversarial Networks for Transient Fault Detection under Various Operation Conditions.

The efficient fault detection (FD) of traction control systems (TCSs) is crucial for ensuring the safe operation of high-speed trains. Transient faults (TFs) can arise due to prolonged operation and harsh environmental conditions, often being masked by background noise, particularly during dynamic operating conditions. Moreover, acquiring a sufficient number of samples across the entire scenario presents a challenging task, resulting in imbalanced data for FD. To address these limitations, an unsupervised transfer learning (TL) method via federated Cycle-Flow adversarial networks (CFANs) is proposed to effectively detect TFs under various operating conditions. Firstly, a CFAN is specifically designed for extracting latent features and reconstructing data in the source domain. Subsequently, a transfer learning framework employing federated CFANs collectively adjusts the modified knowledge resulting from domain alterations. Finally, the designed federated CFANs execute transient FD by constructing residuals in the target domain. The efficacy of the proposed methodology is demonstrated through comparative experiments.

Few-Shot Learning Based on Dimensionally Enhanced Attention and Logit Standardization Self-Distillation

Few-shot learning (FSL) is a challenging problem. Transfer learning methods offer a straightforward and effective solution to FSL by leveraging pre-trained models and generalizing them to new tasks. However, pre-trained models often lack the ability to highlight and emphasize salient features, a gap that attention mechanisms can fill. Unfortunately, existing attention mechanisms encounter issues such as high complexity and incomplete attention information. To address these issues, we propose a dimensionally enhanced attention (DEA) module for FSL. This DEA module introduces minimal additional computational overhead while fully attending to both channel and spatial information. Specifically, the feature map is first decomposed into 1D tensors of varying dimensions using strip pooling. Next, a multi-dimensional collaborative learning strategy is introduced, enabling cross-dimensional information interactions through 1D convolutions with adaptive kernel sizes. Finally, the feature representation is enhanced by calculating attention weights for each dimension using a sigmoid function and weighting the original input accordingly. This approach ensures comprehensive attention to different dimensions of information, effectively characterizing data in various directions. Additionally, we have found that knowledge distillation significantly improves FSL performance. To this end, we implement a logit standardization self-distillation method tailored for FSL. This method addresses the issue of exact logit matching, which arises from the shared temperature in the self-distillation process, by employing logit standardization. We present experimental results on several benchmark datasets where the proposed method yields significant performance improvements.

Classification of Traffic Signs Using Transfer Learning Methods

Transportation refers to a process based on the movement of people or vehicles from one place to another. Sea routes and roads have existed for centuries. They generally play a very important role in people's daily life, trade and industrial activities. Highway, a mode of transportation, is the first preferred mode of transportation worldwide. However, various signs and rules have been set by the authorities to prevent chaos on the highways. Traffic signs are the most important of these rules. In this study, transfer learning models (VGG16, VGG19, Xception and EfficientNet) are used to classify traffic signs using a state-of-art traffic signs dataset (German Traffic Sign Detection Benchmark-GTSDB). Accuracy was used as the classification evaluation criterion. The CNN model designed for the study gave the best result with an accuracy rate of 98% and a model competing with the literature was proposed.

Transfer learning by VGG-16 with convolutional neural network for paddy leaf disease classification

ABSTRACT One of the most recent areas of research in agriculture is the detection along with classification of diseases from plant leaf images. Since rice is the most widely consumed staple food worldwide, it is crucial to increase paddy production’s quality and quantity. In the production of paddy, early detection of pests and diseases at various growth stages is very important. Utilising image processing scheme to spot diseases of agricultural plants will lessen the need for farmers to safeguard agricultural products. The use of VGG-16 along with a Convolutional Neural Network (CNN) for the classification and recognition of paddy leaf diseases is proposed in this work. The rice plant leaves images taken from the kaggle dataset is utilised for the purpose of image acquisition. The Gaussian filter is used in pre-processing. The clustering technique is utilised for the segmentation of the diseased part, the normal part, and the surroundings. The proposed model is then utilised for disease classification. This study classifies 10 categories of paddy images. VGG16, InceptionV3, MobileNetV2, and ResNet-152, among other transfer learning methods, are evaluated and contrasted with the experimental results. The outcome accuracy of the proposed model achieves of nearly 99.9%.

Transfer Learning in Weather Prediction: Why, How, and What Should

Transfer learning (TL) is a popular phrase in deep learning (DL) domain. It is one of the latest artificial intelligence (AI) technologies that has a significant impact on big data analysis. Methods of traditional machine learning (ML) require the availability of an adequate quantity of training data as well as similarity of characteristics among the feature spaces corresponding to training and test data while performing supervised learning tasks. However, in real-life analytical problems, data scarcity often arises. In such scenarios, the TL approach has shown effectiveness in transferring knowledge from the source tasks that had large training data to a target task that has less training data. Basically, in TL, a model that has been trained on one task is essentially applied to a second related (but not exact) task. In this way, the issue of distribution mismatch can also be addressed. TL is not like conventional machine learning algorithms that try to learn each task starting from the beginning. Meteorological research is such an example of big data analysis which often faces the data scarcity issue. The current study addresses the contemporary challenges in weather forecasting that can be solved (or better dealt with) using TL methods. It presents a brief review of earlier research, followed by potential applications of TL algorithms to several key challenges in weather prediction, which includes the prediction of air quality, thunderstorms, precipitation, visibility, and cyclones, among others. Special emphasis is given to high-impact weather (HIW) prediction. These high-impact events are extremely difficult to predict, and they can cause enormous property damage and fatalities around the world. TL techniques have shown advantages in predicting HIW. Various challenging issues in implementing TL technology are then discussed. Finally, we address various prospects associated with TL, propose new research directions, and more importantly mention some concerns for beginners in DL-TL research. An extensive list of references is also provided.

A multi-domain adversarial transfer network for cross domain fault diagnosis under imbalanced data

In the intelligent fault diagnosis of rolling bearings, transfer learning methods extend the applicability of models to diverse working scenarios. However, real-world scenarios often suffer from data imbalance, which reduces diagnostic accuracy. To address this issue, this paper proposes a multi-domain adversarial transfer (MDAT) framework to enhance cross-domain fault diagnosis accuracy for rolling bearings under imbalanced data conditions. First, an enhanced information generation method is introduced to produce realistic and useable synthetic data to mitigate data imbalance. Subsequently, an adversarial multi-domain adaptation module is designed to learn invariant features across multiple domains. Finally, a domain reweighting method is proposed to improve domain alignment and enhance domain confusion. The effectiveness of the proposed method was validated through two case studies on rolling bearing fault diagnosis. The results demonstrated that MDAT achieved cross-domain diagnosis accuracies of 89.2% and 99.0% under imbalanced data conditions, confirming the effectiveness and superiority of the MDAT framework.

Fused feature extract method for Φ-OTDR event recognition based on VGGish transfer learning

Thanks to the development of artificial intelligence algorithms, the event recognition of distributed optical fiber sensing systems has achieved high classification accuracy on many deep learning models. However, the large-scale samples required for the deep learning networks are difficult to collect for the optical fiber vibration sensing systems in actual scenarios. An overfitting problem due to insufficient data in the network training process will reduce the classification accuracy. In this paper, we propose a fused feature extract method suitable for the small dataset of Φ-OTDR systems. The high-dimensional features of signals in the frequency domain are extracted by a transfer learning method based on the VGGish framework. Combined with the characteristics of 12 different acquisition points in the space, the spatial distribution characteristics of the signal can be reflected. Fused with the spatial and temporal features, the features undergo a sample feature correction algorithm and are used in a SVM classifier for event recognition. Experimental results show that the VGGish, a pre-trained convolutional network for audio classification, can extract the knowledge features of Φ-OTDR vibration signals more efficiently. The recognition accuracy of six types of intrusion events can reach 95.0% through the corrected multi-domain features when only 960 samples are used as the training set. The accuracy is 17.7% higher than that of the single channel trained on VGGish without fine-tuning. Compared to other CNNs, such as ResNet, the feature extract method proposed can improve the accuracy by at least 4.9% on the same dataset.

A Comparative Study on Thyroid Nodule Classification Using Transfer Learning Methods

A Comparative Study on Thyroid Nodule Classification Using Transfer Learning Methods

An Computational Framework for Enhanced Early Detection of Diabetic Retinopathy with Transfer Learning Methods

Easier early spotting of diabetes retinopathy is important for getting patients help quickly and improving their results. In this research, we used transfer learning techniques to make deep learning models better at automatically finding diabetic retinopathy. Firstly, we used methods for "data augmentation" to make our training sample more diverse. This made our models better at generalization. After this proposed method used fine-tuned the hyperparameters of the transfer learning models. Four transfer learning models caught our attention as MobileNetV2, EfficientNet, VGG16, and a Hybrid Ensemble (DenseNet + Inception). The method focused on the DenseNet model because it had good results in past research and could clearly pick out complex details in medical Images. When combined with data enhancement and fine-tuning, our experiments showed that the DenseNet model worked better than the others in terms of both accuracy and speed. This model, DenseNet, had an amazing success rate of 88.81%, showing that it could be a useful tool for finding diabetic retinopathy early on. The work shows how important transfer learning is in medical image analysis and how to use deep learning models most effectively for finding diabetic retinopathy early. Using advanced machine learning methods to help find and fix diabetic retinopathy will be studied more in the future.

A Cross-Resolution Surface Net Radiative Inversion Based on Transfer Learning Methods

Net radiation (Rn) is a key component of the Earth’s energy balance. With the rise of deep learning technology, remote sensing technology has made significant progress in the acquisition of large-scale surface parameters. However, the generally low spatial resolution of net radiation data and the relative scarcity of surface flux site data at home and abroad limit the potential of deep learning methods in constructing high spatial resolution net radiation models. To address this challenge, this study proposes an innovative approach of a multi-scale transfer learning framework, which assumes that composite models at different spatial scales are similar in structure and parameters, thus enabling the training of accurate high-resolution models using fewer samples. In this study, the Heihe River Basin was taken as the study area and the Rn products of the Global Land Surface Satellite (GLASS) were selected as the target for coarse model training. Based on the dense convolutional network (DenseNet) architecture, 25 deep learning models were constructed to learn the spatial and temporal distribution patterns of GLASS Rn products by combining multi-source data, and a 5 km coarse resolution net radiation model was trained. Subsequently, the parameters of the pre-trained coarse-resolution model were fine-tuned with a small amount of measured ground station data to achieve the transfer from the 5 km coarse-resolution model to the 1 km high-resolution model, and a daily high-resolution net radiation model with 1 km resolution for the Heihe River Basin was finally constructed. The results showed that the bias, R2, and RMSE of the high-resolution net radiation model obtained by transfer learning were 0.184 W/m2, 0.924, and 24.29 W/m2, respectively, which was better than those of the GLASS Rn products. The predicted values were highly correlated with the measured values at the stations and the fitted curves were closer to the measured values at the stations than those of the GLASS Rn products, which further demonstrated that the transfer learning method could capture the soil moisture and temporal variation of net radiation. Finally, the model was used to generate 1 km daily net radiation products for the Heihe River Basin in 2020. This study provides new perspectives and methods for future large-scale and long-time-series estimations of surface net radiation.

Implementation Transfer Learning on Convolutional Neural Network for Tubercolosis Classification

Tuberculosis (TB) is an infectious disease that can have serious effects on the lungs and is among the top 10 causes of death worldwide. This disease is caused by the transmission of Mycobacterium tuberculosis bacteria through the air when coughing or sneezing. Without treatment, pulmonary tuberculosis can result in permanent lung damage and can be life-threatening. Accurate and early diagnosis is crucial for effective treatment and control of the disease.The challenge lies in the accurate classification of tuberculosis from lung images, which is essential for timely diagnosis and treatment. Traditional diagnostic methods can be time-consuming and sometimes lack precision. To address this issue, this research aims to achieve high accuracy in classifying tuberculosis using the Convolutional Neural Network (CNN) algorithm through transfer learning methods. By utilizing visual images of tuberculosis-affected and normal lungs, we propose a solution that leverages advanced deep learning techniques to enhance diagnostic accuracy. This approach not only expedites the diagnostic process but also improves the reliability of tuberculosis detection, ultimately contributing to better patient outcomes and more effective disease management. The dataset applied consists of two labels: tuberculosis and normal. This dataset contains 4200 lung images of individuals with tuberculosis and normal lungs. By applying the transfer learning method, Transfer learning is a machine learning method where a pre-trained model is used as the starting point for a new, related task. it was found that the ResNet50 model achieved the highest accuracy at 99%, followed by InceptionV3 at 97%, and lastly, DenseNet121 at 91%.

A living lab for Structural Health Monitoring at the Nibelungen Bridge Worms for Transfer Learning of Structural Dynamics

The Nibelungen Bridge in Worms, Germany has been selected as a national demonstration structure for advanced non-destructive testing (NDT) and structural health monitoring concepts to extend the lifetime of civil structures and to optimize O&M actions. Parts of the research that involves this bridge as a demonstrator belong to the focus area program SPP100+. In this program, the bridges SHM System has been extended and combined with an additional setup of vibration sensors. The used digital smart sensor with pre-processing functions, the arrangement of the sensors at the structure and additional edge computing capability allows the investigation of transfer learning and other methods directly into the real structure. The living lab with seven triaxial accelerometers can be reconfigured in real-time and adjusted to the needs of AI models for classification. The comparison with the existing conventional SHM sensors has been made possible by hardware synchronization to the existing SHM System and collocating sensors at similar positions, so that a hardware exchange can be an investigated use-case for the transfer learning. During idle times, the system collects vibration data like a conventional SHM system. Keywords: Structural Health Monitoring (SHM), Data acquisition, Sensor technology.

State of Charge Estimation Method of Energy Storage Battery Based on Multiple Incremental Features

Accurately estimating the state of charge (SOC) is crucial for energy storage battery management systems as it ensures battery performance and extends lifespan. However, existing deep learning-based methods often overlook the dynamic process information during battery charging and discharging, which compromises the accuracy of SOC estimation. To address this limitation, this paper proposes a novel SOC estimation method. First, we employ differential processing on the collected voltage, current, and temperature data to capture dynamic feature changes. Next, all features are normalized to ensure they are on the same scale. Finally, the processed data is divided into sliding windows and input into the TCN-BiLSTM-Attention Net (TBANet) model for SOC estimation. The results show that compared with traditional deep learning based SOC estimation methods, adding incremental features to TBANet improves the estimation accuracy by 15.8%. The average absolute error and root mean square error of the experimental results are 0.72% and 0.91%, respectively. In addition, this approach adopts transfer learning methods to verify the strong adaptability of the proposed method on different datasets, which highlights the robustness of TBANet and its potential for wide applicability in real-world scenarios.

A Systematic Literature Review on Leaf Disease Recognition Using Computer Vision and Deep Learning Approach

Background: Plant diseases affect agricultural output, quality and profitability, making them serious obstacles for agriculture. It is essential to detect diseases early in order to reduce losses while retaining sustainable practices. Plant disease detection has benefited greatly from the use of computer vision and deep learning in recent years because of their outstanding precision and computing capability. Objective: In this paper, we intend to investigate the role of deep learning in computer vision for plant disease detection while looking into how these techniques address complex disease identification problems. A variety of deep learning architectures were reviewed, and the contribution of frameworks such as Tensorflow, Keras, Caffe and PyTorch to the researchers' model construction was studied as well. Additionally, the usage of open repositories such as PlantVillage and Kaggle along with the customized datasets were discussed. Methods: We gathered the most recent developments in deep learning techniques for leaf disease detection through a systematic literature review of research papers published over the past decade, using reputable academic databases like Scopus and Web of Science, following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) method for guidance. Results: This study finds that researchers consistently enhance existing deep learning architectures to improve prediction accuracy in plant disease detection, often by introducing novel architectures and employing transfer learning methods. Frameworks like TensorFlow, Keras, Caffe, and PyTorch are widely favored for their efficiency in development. Additionally, most studies opt for public datasets such as PlantVillage, Kaggle, and ImageNet, which offer an abundance of labelled data for training and testing deep learning models. Conclusion: While no singular ‘best' model emerges, the adaptability of deep learning and computer vision demonstrates the dynamic nature of plant disease recognition area, and this paper provides a comprehensive overview of deep learning's transformative impact on plant disease recognition by bringing together information from different studies. Keywords: Deep learning, Computer vision, Plant disease, Systematic literature review

Machining quality prediction of multi-feature parts using integrated multi-source domain dynamic adaptive transfer learning

Machining quality prediction of multi-feature parts has been a challenging problem because of small dataset and inconsistent quality data distribution with respect to each machining feature. Transfer learning that leverages knowledge of one task and can be repurposed on another task seems a good solution for this purpose. However, traditional transfer learning typically has a single source domain and a target domain, which limits its applications in multi-source scenarios (e.g., multi-feature). To solve this issue, this paper proposes a novel integrated multi-source domain dynamic adaptive transfer learning (IMD-DATL) framework for machining quality prediction of multi-feature part machining systems. Specifically, a domain-sample similarity double matching multi-source domain integration method is designed to construct the integration knowledge transfer from multiple source domains to the target domain. A residual feature extraction network based on sample entropy-dynamic channel double-layer attention structure and a fine-grained transferable feature attention module are designed. These three attentions are used to improve the feature learning ability and the adaptation level to the predicted object in the three dimensions of sample, channel and data feature. Finally, multiple sets of comparative experiments in thin-walled part machining systems confirm the effectiveness and superiority of the proposed method for cross-domain quality prediction. Compared with other traditional transfer learning methods, the MAE, RMSE and Score on average of this method are increased by 5.47 %, 4.59 % and 4.84 %, respectively, compared with other multi-source domain adaptation methods, the MAE, RMSE and Score on average of this method are increased by 7.13 %, 7.37 % and 6.52 %, respectively.

Time series modeling and forecasting of epidemic spreading processes using deep transfer learning

Traditional data-driven methods for modeling and predicting epidemic spreading typically operate in an independent and identically distributed setting. However, epidemic spreading on complex networks exhibits significant heterogeneity across different phases, regions, and viruses, indicating that epidemic time series may not be independent and identically distributed due to temporal and spatial variations. In this article, a novel deep transfer learning method integrating convolutional neural networks (CNNs) and bi-directional long short-term memory (BiLSTM) networks is proposed to model and forecast epidemics with heterogeneous data. The proposed method combines a CNN-based layer for local feature extraction, a BiLSTM-based layer for temporal analysis, and a fully connected layer for prediction, and employs transfer learning to enhance the generalization ability of the CNN-BiLSTM model. To improve prediction performance, hyperparameter tuning is conducted using particle swarm optimization during model training. Finally, we adopt the proposed approach to characterize the spatio-temporal spreading dynamics of COVID-19 and infer the pathological heterogeneity among epidemics of severe acute respiratory syndrome (SARS), influenza A (H1N1), and COVID-19. The comprehensive results demonstrate the effectiveness of the proposed approach in exploring the spatiotemporal variations in the spread of epidemics and characterizing the epidemiological features of different viruses. Moreover, the proposed method can significantly reduce modeling and predicting errors in epidemic spread to some extent.

Transfer learning for cross-modal demand prediction of bike-share and public transit

The urban transportation system is a combination of multiple transport modes, and the interdependencies across those modes exist. This means that the travel demand across different travel modes could be correlated as one mode may receive/create demand from/for another mode, not to mention natural correlations between different demand time series due to general demand flow patterns across the network. It is expected that cross-modal ripple effects will become more prevalent with Mobility as a Service. Therefore, by propagating demand data across modes, a better demand prediction could be obtained. To this end, this study explores various transfer learning strategies and machine learning models for cross-modal demand prediction. The trip data of bike-share, metro, and taxi are processed as the station-level passenger flows, and then the proposed prediction method is tested in the large-scale case studies of Nanjing and Chicago. The results suggest that prediction models with transfer learning perform better than unimodal prediction models. Fine-tuning without freezing strategy performs the best among all transfer learning strategies, and the split-brain strategy can handle the data missing problem. Furthermore, the 3-layer stacked Long Short-Term Memory model performs particularly well in cross-modal demand prediction. These results verify our deep transfer learning method’s forecasting improvement over existing benchmarks and demonstrate the good transferability for cross-modal demand prediction in multiple cities.