Large Number Of Training Samples Research Articles

Hyperspectral Image Restoration With Self-Supervised Learning: A Two-Stage Training Approach

Hyperspectral image (HSI) denoising is a crucial preprocessing task to improve the performance of the subsequent HSI interpretation and applications. With recent progress in deep learning, HSI denoising methods based on deep neural networks have attracted increasing interest and achieved the state-of-the-art performance. Nevertheless, most of these methods are based on network structures originally developed for grayscale and color images and require a change of network structure to be applicable to HSIs. The new network architectures often lead to complicated models and limited flexibility, which, in turn, result in difficulty in learning and demand of a large number of training samples. In this article, we propose an innovative two-stage learning method including pretraining and fine-tuning procedures. In the first stage, a denoising convolutional neural network can be pretrained with pairs of corrupted and clean images. In the second stage, the pretrained network is fine-tuned via a self-supervised learning strategy to capture the spectral correlation in HSIs. The training pairs in the second stage are constructed from the neighboring band images in the target noisy HSI, leading to a novel idea of embedding spectral information into denoiser through the target image, rather than the change of the network architecture. This model has strong adaptability such that many image denoising networks can be easily adopted for HSIs, while the external hyperspectral training set is optional but not mandatory. Experimental results show that our method has competitive performance compared with the state-of-the-art approaches.

Data Augmentation via Randomized Wavelet Expansion and Its Application in Few-Shot Fault Diagnosis of Aviation Hydraulic Pumps

In general, deep learning-based fault diagnosis methods need a large number of training samples, which are often not available in real applications. Aiming at this problem, this article develops a new data augmentation method, i.e., randomized wavelet expansion (RWE), to generate a set of synthesis samples that share similar characteristics with the original sample. The first key point is that the amplitudes of wavelet coefficients at a randomly selected frequency band are enlarged through random expansion. Another key point is that the synthesis samples are processed to have the same mean values and standard deviations as their corresponding original sample. Afterward, the synthesis samples are used as the training dataset to train a deep convolutional neural network (CNN) for implementing the few-shot fault diagnosis of aviation hydraulic pumps. Finally, the performance has been validated through a series of experiments.

Radio Frequency Interference Signature Detection in Radar Remote Sensing Image Using Semantic Cognition Enhancement Network

Radio frequency interference (RFI) is a significant threat to accurate microwave remote sensing. The RFI signals manifest themselves in unpredictable locations and patterns in the image, which will cause measurement distortion, image degradation, or even lead to wrong retrievals of the geophysical parameters. Accurate detection of RFI artifacts is a prerequisite step to preserve the overall quality of remote sensing quality. In this paper, a semantic cognitive enhancement network for RFI signature detection is proposed. It employs an encoder-decoder architecture, which incorporates the atrous spatial pyramid pooling, Depthwise convolution, and self-attentional mechanism. Rather than detecting the existence of RFI artifacts for an entire image, the proposed scheme can realize RFI recognition in a pixel-wise manner without setting predefined thresholds. Extensive experimental results on diverse scenarios in Sentinel-1 images with various RFI types are provided, which demonstrates robust detection performance for both strong and weak interference without requiring a large number of training samples.

Robust Matching for SAR and Optical Images Using Multiscale Convolutional Gradient Features

Image matching is a key preprocessing step for the integrated application of synthetic aperture radar (SAR) and optical images. Due to significant nonlinear intensity differences between such images, automatic matching for them is still quite challenging. Recently, structure features have been effectively applied to SAR-to-optical image matching because of their robustness to nonlinear intensity differences. However, structure features designed by handcraft are limited to achieve further improvement. Accordingly, this letter employs the deep learning technique to refine structure features for improving image matching. First, we extract multiorientated gradient features to depict the structure properties of images. Then, a shallow pseudo-Siamese network is built to convolve the gradient feature maps in a multiscale manner, which produces the multiscale convolutional gradient features (MCGFs). Finally, MCGF is used to achieve image matching by a fast template scheme. MCGF can capture finer common features between SAR and optical images than traditional handcrafted structure features. Moreover, it also can overcome some limitations of current matching methods based on deep learning, which requires solving a huge number of model parameters by a large number of training samples. Two sets of SAR and optical images with different resolutions are used to evaluate the matching performance of MCGF. The experimental results show its advantage over other state-of-the-art methods.

A dual-scale siamese densely connected network with MRF for SAR image classification

ABSTRACT Recently, Convolutional Neural Network (CNN) has achieved some success in synthetic aperture radar (SAR) image classification. This outstanding performance mainly depends on a large number of training samples, but achieved satisfactory classification results with limited training samples remains a challenge. To address this problem, we propose a dual-scale Siamese densely connected network with Markov Random Fields (DS-SDCNet-MRF) for single-polarization SAR image classification. First, the Siamese densely connected network (SDCNet) is proposed to fully extract discriminative features under limited samples. Then, the proposed DS-SDCNet is constructed with two SDCNets of different scales to produce complementary classification results. Among them, the large-scale SDCNet has a better classification result in the homogeneous region, while the small-scale SDCNet tends to provide good detail preservation. Finally, an improved MRF model which combined the category probability information is proposed to further improve the classification performance. Experimental results on simulated and real single-polarization SAR data demonstrate that the proposed method achieves more encouraging classification performance than the current state-of-the-art classification methods with limited samples.

A Few-shot Learning Method for Aging Diagnosis of Oil-paper Insulation by Raman Spectroscopy Based on Graph Theory

Raman spectroscopy is an effective method for diagnosing the aging state of transformer oil-paper insulation. Aiming at overcoming the drawback of previous methods that require a large number of training samples, this research proposes a method for aging diagnosis of oil-paper insulation that can be used for few-shot learning. In this study, different Raman spectroscopy samples of aging insulating oil are obtained by accelerated thermal aging experiments. This paper uses identity matrix as node feature, calculates the connection weights between them by Gaussian kernel function, and constructs the graph structure of Raman spectroscopy library. Using graph convolutional neural network (GCN) based on the graph structure, a diagnosis model is established. The diagnostic accuracy rate is as high as 95.5%. This method provides a theoretical basis for establishing a new structured Raman spectroscopy library and assessing the aging state of transformer oil-paper insulation with a small number of training samples.

Semi-Automatic Generation of Training Samples for Detecting Renewable Energy Plants in High-Resolution Aerial Images

Deep learning (DL)—in particular convolutional neural networks (CNN)—methods are widely spread in object detection and recognition of remote sensing images. In the domain of DL, there is a need for large numbers of training samples. These samples are mostly generated based on manual identification. Identifying and labelling these objects is very time-consuming. The developed approach proposes a partially automated procedure for the sample creation and avoids manual labelling of rooftop photovoltaic (PV) systems. By combining address data of existing rooftop PV systems from the German Plant Register, the Georeferenced Address Data and the Official House Surroundings Germany, a partially automated generation of training samples is achieved. Using a selection of 100,000 automatically generated samples, a network using a RetinaNet-based architecture combining ResNet101, a feature pyramid network, a classification and a regression network is trained, applied on a large area and post-filtered by intersection with additional automatically identified locations of existing rooftop PV systems. Based on a proof-of-concept application, a second network is trained with the filtered selection of approximately 51,000 training samples. In two independent test applications using high-resolution aerial images of Saarland in Germany, buildings with PV systems are detected with a precision of at least 92.77 and a recall of 84.47.

Improving the Accuracy of Land Cover Mapping by Distributing Training Samples

High-quality training samples are essential for accurate land cover classification. Due to the difficulties in collecting a large number of training samples, it is of great significance to collect a high-quality sample dataset with a limited sample size but effective sample distribution. In this paper, we proposed an object-oriented sampling approach by segmenting image blocks expanded from systematically distributed seeds (object-oriented sampling approach) and carried out a rigorous comparison of seven sampling strategies, including random sampling, systematic sampling, stratified sampling (stratified sampling with the strata of land cover classes based on classification product, Latin hypercube sampling, and spatial Latin hypercube sampling), object-oriented sampling, and manual sampling, to explore the impact of training sample distribution on the accuracy of land cover classification when the samples are limited. Five study areas from different climate zones were selected along the China–Mongolia border. Our research identified the proposed object-oriented sampling approach as the first-choice sampling strategy in collecting training samples. This approach improved the diversity and completeness of the training sample set. Stratified sampling with strata defined by the combination of different attributes and stratified sampling with the strata of land cover classes had their limitations, and they performed well in specific situations when we have enough prior knowledge or high-accuracy product. Manual sampling was greatly influenced by the experience of interpreters. All these sampling strategies mentioned above outperformed random sampling and systematic sampling in this study. The results indicate that the sampling strategies of training datasets do have great impacts on the land cover classification accuracies when the sample size is limited. This paper will provide guidance for efficient training sample collection to increase classification accuracies.

Bearing Prognostics: An Instance-Based Learning Approach with Feature Engineering, Data Augmentation, and Similarity Evaluation

We propose an instance-based learning approach with data augmentation and similarity evaluation to estimate the remaining useful life (RUL) of a mechanical component for health management. The publicly available PRONOSTIA datasets, which provide accelerated degradation test data for bearings, are used in our study. The challenges with the datasets include a very limited number of run-to-failure examples, no failure mode information, and a wide range of bearing life spans. Without a large number of training samples, feature engineering is necessary. Principal component analysis is applied to the spectrogram of vibration signals to obtain prognostic feature sequences. A data augmentation strategy is developed to generate synthetic prognostic feature sequences using learning instances. Subsequently, similarities between the test and learning instances can be assessed using a root mean squared (RMS) difference measure. Finally, an ensemble method is developed to aggregate the RUL estimates based on multiple similar prognostic feature sequences. The proposed approach demonstrates comparable performance with published solutions in the literature. It serves as an alternative method for solving the RUL estimation problem.

Improved Visual Recognition Memory Model Based on Grid Cells for Face Recognition.

Traditional facial recognition methods depend on a large number of training samples due to the massive turning of synaptic weights for low-level feature extractions. In prior work, a brain-inspired model of visual recognition memory suggested that grid cells encode translation saccadic eye movement vectors between salient stimulus features. With a small training set for each recognition type, the relative positions among the selected features for each image were represented using grid and feature label cells in Hebbian learning. However, this model is suitable only for the recognition of familiar faces, objects, and scenes. The model's performance for a given face with unfamiliar facial expressions was unsatisfactory. In this study, an improved computational model via grid cells for facial recognition was proposed. Here, the initial hypothesis about stimulus identity was obtained using the histograms of oriented gradients (HOG) algorithm. The HOG descriptors effectively captured the sample edge or gradient structure features. Thus, most test samples were successfully recognized within three saccades. Moreover, the probability of a false hypothesis and the average fixations for successful recognition were reduced. Compared with other neural network models, such as convolutional neural networks and deep belief networks, the proposed method shows the best performance with only one training sample for each face. Moreover, it is robust against image occlusion and size variance or scaling. Our results may give insight for efficient recognition with small training samples based on neural networks.

Galaxy morphology classification using automated machine learning

In this paper, we apply five different machine learning algorithms to classify samples into four categories — spirals, ellipticals, mergers and stars (don’t know) using data from the Sloan Digital Sky Survey to assess the feasibility of using machine learning methods for future surveys. Classifying mergers as a separate class poses a challenge as this category is easily confused with both ellipticals and spirals, and as a result, most previous studies have not included mergers as a distinct morphological class. The dataset is highly imbalanced with the number of ellipticals/spirals being much larger than the number of stars/mergers, and this is another challenge we aim to address. Starting with 62 features, we perform principal component analysis and use the 25 most significant principal components as inputs to the machine learning models. We compare our results with the Galaxy Zoo labels and obtain an overall test accuracy of 98.2% and 97.5% using Artificial Neural Network and ExtraTrees respectively. However, ExtraTrees outperforms Neural Network in classifying mergers and stars. We also perform a parameter sensitivity test to compare the relative importance of different categories of features on the model’s performance. Finally, we address the class imbalance problem and examine the effects of different sampling strategies. Our results show that the use of a balanced dataset with a large number of training samples leads to high recall values for the minority classes, and that oversampling methods lead to better performance than undersampling techniques.

Small CSI Samples-Based Activity Recognition: A Deep Learning Approach Using Multidimensional Features

With the emergence of tools for extracting CSI data from commercial WiFi devices, CSI-based device-free activity recognition technology has developed rapidly and has been widely used in security monitoring, smart home, medical monitoring, and other fields. However, the existing CSI-based activity recognition algorithms need a large number of training samples to obtain the ideal recognition accuracy. To solve the problem, an attention-based bidirectional LSTM method using multidimensional features (called MF-ABLSTM method) is proposed. In this method, the signal preprocessing and continuous wavelet transform algorithms are used to construct time-frequency matrix, the sample entropy is used to characterize the statistical feature of CSI amplitudes, the energy difference at a fixed time interval is used to characterize the time-domain feature of activities, and the energy distribution of different frequency components is used to characterize the frequency-domain feature of activities. By expanding the training samples with the proposed tensor prediction algorithm, the accurate activity recognition can be realized with only a few samples. A large number of experiments verify the good performance of MF-ABLSTM method.

Object-Level Remote Sensing Image Augmentation Using U-Net-Based Generative Adversarial Networks

With the continuous development of deep learning in computer vision, semantic segmentation technology is constantly employed for processing remote sensing images. For instance, it is a key technology to automatically mark important objects such as ships or port land from port area remote sensing images. However, the existing supervised semantic segmentation model based on deep learning requires a large number of training samples. Otherwise, it will not be able to correctly learn the characteristics of the target objects, which results in the poor performance or even failure of semantic segmentation task. Since the target objects such as ships may move from time to time, it is nontrivial to collect enough samples to achieve satisfactory segmentation performance. And this severely hinders the performance improvement of most of existing augmentation methods. To tackle this problem, in this paper, we propose an object-level remote sensing image augmentation approach based on leveraging the U-Net-based generative adversarial networks. Specifically, our proposed approach consists two components including the semantic tag image generator and the U-Net GAN-based translator. To evaluate the effectiveness of the proposed approach, comprehensive experiments are conducted on a public dataset HRSC2016. State-of-the-art generative models, DCGAN, WGAN, and CycleGAN, are selected as baselines. According to the experimental results, our proposed approach significantly outperforms the baselines in terms of not only drawing the outlines of target objects but also capturing their meaningful details.

Hybrid attention mechanism for few‐shot relational learning of knowledge graphs

Few-shot knowledge graph (KG) reasoning is the main focus in the field of knowledge graph reasoning. In order to expand the application fields of the knowledge graph, a large number of studies are based on a large number of training samples. However, we have learnt that there are actually many missing relationships or entities in the knowledge graph, and in most cases, there are not many training instances when implementing new relationships. To tackle it, in this study, the authors aim to predict a new entity given few reference instances, even only one training instance. A few-shot learning framework based on a hybrid attention mechanism is proposed. The framework employs traditional embedding models to extract knowledge, and uses an attenuated attention network and a self-attention mechanism to obtain the hidden attributes of entities. Thus, it can learn a matching metric by considering both the learnt embeddings and one-hop graph structures. The experimental results present that the model has achieved significant performance improvements on the NELL-One and Wiki-One datasets.

An image-level weakly supervised segmentation method for No-service rail surface defect with size prior

Surface defect segmentation of no-service rail is important for its quality assessment. Several deep learning methods have obtained good performances based on computer vision testing. However, these methods all require a large number of training samples with pixel-level labels. Marking pixel-level labels is a difficult task – time-consuming, labor-consuming, and professionals. To address the issue, a novel image-level weakly supervised segmentation formulation is proposed for no-service rail surface defects. These defects are decomposed into three sub-categories (strip-shaped, spot-shaped, block-shaped) according to the size prior information (area and shape). Then, a method is presented with a pooling combination module. The pooling combination module makes full use of the size attributes of the sub-category by utilizing different pooling functions for different sub-categories. Experimental results demonstrate that our method is effective and outperforms the state-of-the-art methods.Index terms - Surface defect segmentation; no-service rail; image-level weakly supervised; size prior information; pooling combination module.

Nonlinear variational Bayesian Student’s-t mixture regression and inferential sensor application with semisupervised data

Gaussian mixture models have been widely applied to predict those significant but hard-to-measure quality variables in industrial processes. However, outliers that result from disturbances of noise, and incorrect instrument readings, will deteriorate the prediction accuracy of Gaussian mixture models. Student’s-t mixture regression and its variants have been proposed for robust inferential sensor development. Nevertheless, for those Student’s-t mixture regression-based inferential sensors, the functional dependency between the quality and process variables in each component is assumed to be linear. Such configuration in practice might be incapable of adequately modeling the complicated nonlinearity between the quality and process variables. In addition, large amounts of unlabeled samples are easily to be obtained and informative for model performance improvement, however, the computations are significantly increased as well. To get over these issues, this paper proposes a novel semisupervised nonlinear variational Bayesian Student’s-t mixture regression model, in which the nonlinear component model can effectively capture the non-linearity between the quality and process variables. For tackling the low training efficiency issue caused by large number of training samples, the parallel computing strategy is introduced to complete model training. A numerical example and two real industrial processes are supplied to validate the effectiveness of the proposed method.

Multi-View Feature Construction Using Genetic Programming for Rolling Bearing Fault Diagnosis [Application Notes

Rolling bearing fault diagnosis is an important task in mechanical engineering. Existing methods have several limitations, such as requiring domain knowledge and a large number of training samples. To address these limitations, this paper proposes a new diagnosis approach, i.e., multiview feature construction based on genetic programming with the idea of ensemble learning (MFCGPE), to automatically construct high-level features from multiple views and build an effective ensemble for identifying different fault types using a small number of training samples. The MFCGPE approach uses a new program structure to automatically construct a flexible number of features from every single view. A new fitness function based on accuracy and distance is developed in MFCGPE to improve the discriminability of the constructed features. To further improve the generalization performance, an ensemble of classifiers based on k-nearest neighbor is created by using the constructed features from every single view. Three bearing datasets and 19 competitive methods are used to validate the effectiveness of the new approach. The results show that MFCGPE achieves higher diagnostic accuracy than all the compared methods on the three datasets with a small number of training samples.

Multiple-view D2NNs array: realizing robust 3D object recognition.

As an optical-based classifier of the physical neural network, the independent diffractive deep neural network (D2NN) can be utilized to learn the single-view spatial featured mapping between the input lightfields and the truth labels by preprocessing a large number of training samples. However, it is still not enough to approach or even reach a satisfactory classification accuracy on three-dimensional (3D) targets owing to already losing lots of effective lightfield information on other view fields. This Letter presents a multiple-view D2NNs array (MDA) scheme that provides a significant inference improvement compared with individual D2NN or Res-D2NN by constructing a different complementary mechanism and then merging all base learners of distinct views on an electronic computer. Furthermore, a robust multiple-view D2NNs array (r-MDA) framework is demonstrated to resist the redundant spatial features of invalid lightfields due to severe optical disturbances.

Red wine quality prediction through active learning

General machine learning requires a large number of training samples, meaning a lot of manpower and material resources. To solve this problem, active learning and its algorithms are often used. Active learning is a form of semi-supervised machine learning where the algorithm can choose which data it wants to learn from and then use the smallest and most effective labeled data set to make predictions or classifications. This article will show an example of using active learning to predict red wine quality. The predictive modeling approach the author chose was the K-Nearest Neighbor and the active learning algorithm was ranked batch-mode sampling. By observing the learning curve, the author found that generally the prediction accuracy of active learning would increase as the number of iterations increased. The author compared the experiment with another case using classic iris flower data set, and concluded that the prediction accuracy of active learning for different data sets depends on many factors, such as the correlation between the independent and dependant variables, the size of the data set and the number of iterations.

SPATIAL RESLUTION SENSITIVITY ANALYSIS OF CLASSIFCIATION OF SENTINEL-2 IMAGES BY PRE-TRAINED DEEP MODELS FROM BIG EARTH NET DATABASE

Abstract. Classifying and monitoring different vegetation types is important for forest management, food resources, and assessing the potential impacts of climate change. In this regard, several methods have been developed to study them using remote sensing data, and with the advent of neural networks, new methods are being proposed, especially in the field of automatic land use classification. In this research, multispectral Sentinel-2 satellite image has been used due to having spectral information and different spatial resolution for classifying plant species. Deep learning models have the ability to learn and recognize different features of images, but require a large number of training samples, so we used pre-trained ResNet networks with depths of 50, 101 and 152 layers, that trained with BigEarthNet dataset. The main purpose of this study is to evaluate the sensitivity of ResNet networks to spatial resolution. Results show that ResNet 101 was more stable than other networks, and the Resent 50 with an overall accuracy of 76.2 has the highest accuracy at a resolution of 20 meters.