Lack Of Training Samples Research Articles

Data Augmentation for EEG-Based Emotion Recognition Using Generative Adversarial Networks.

One of the greatest limitations in the field of EEG-based emotion recognition is the lack of training samples, which makes it difficult to establish effective models for emotion recognition. Inspired by the excellent achievements of generative models in image processing, we propose a data augmentation model named VAE-D2GAN for EEG-based emotion recognition using a generative adversarial network. EEG features representing different emotions are extracted as topological maps of differential entropy (DE) under five classical frequency bands. The proposed model is designed to learn the distributions of these features for real EEG signals and generate artificial samples for training. The variational auto-encoder (VAE) architecture can learn the spatial distribution of the actual data through a latent vector, and is introduced into the dual discriminator GAN to improve the diversity of the generated artificial samples. To evaluate the performance of this model, we conduct a systematic test on two public emotion EEG datasets, the SEED and the SEED-IV. The obtained recognition accuracy of the method using data augmentation shows as 92.5 and 82.3%, respectively, on the SEED and SEED-IV datasets, which is 1.5 and 3.5% higher than that of methods without using data augmentation. The experimental results show that the artificial samples generated by our model can effectively enhance the performance of the EEG-based emotion recognition.

Intricate Face Recognition Based On Virtual Sample Generation

This paper mainly focuses on the problem of low recognition rate and making deep hidden features unable to learn caused by the lack of training samples in intricate facial recognition, a method of virtual image sample generation is proposed. The starGAN network based on style migration is used to generate the image that does not exist in the original database, and multiple images with different facial attributes are added to a single sample, combine the obtained virtual samples with the original database to build a larger face database for face recognition. Experiments on the CelebA face database show that the proposed method can effectively improve the face recognition rate in complex situations.

Graph-Based Dimensionality Reduction for Hyperspectral Imagery: A Review

Hyperspectral image (HSI) contains a wealth of spectral information, which makes fine classification of ground objects possible. In the meanwhile, overly redundant information in HSI brings many challenges. Specifically, the lack of training samples and the high computational cost are the inevitable obstacles in the design of classifier. In order to solve these problems, dimensionality reduction is usually adopted. Recently, graph-based dimensionality reduction has become a hot topic. In this paper, the graph-based methods for HSI dimensionality reduction are summarized from the following aspects. 1) The traditional graph-based methods employ Euclidean distance to explore the local information of samples in spectral feature space. 2) The dimensionality-reduction methods based on sparse or collaborative representation regard the sparse or collaborative coefficients as graph weights to effectively reduce reconstruction errors and represent most important information of HSI in the dictionary. 3) Improved methods based on sparse or collaborative graph have made great progress by considering global low-rank information, local intra-class information and spatial information. In order to compare typical techniques, three real HSI datasets were used to carry out relevant experiments, and then the experimental results were analysed and discussed. Finally, the future development of this research field is prospected.

Residual networks with multi-attention mechanism for hyperspectral image classification

Residual network which can effectively overcome gradient disappearance in convolutional neural networks has been successfully applied to hyperspectral classification. Yet simply accumulating residual units does not improve the model performance. Besides, the lack of training samples seriously restricts the hyperspectral classification accuracy. In attempt to solve the abovementioned problems, residual networks with multi-attention mechanism (RNMA) is proposed. The spectral-spatial attention module (SSAM) in accordance with the structure characteristics of hyperspectral images is introduced into the RNMA method. The SSAM is embedded after the input which can strengthen bands relevant to ground object and spatial salient regions of ground object while weaken less informative features. Meanwhile, channel attention module is introduced into residual network unit to form residual network unit with channel attention module (RUCA). Based on RUCA, the features of the spectral dimension and the spatial dimension are sequentially extracted. Besides, sample augmentation method with spectral and spatial constraints is applied to expand sample size. Experiments results on three hyperspectral datasets show that the RNMA model outperforms than other methods in classification accuracy. The RNMA model achieved superior classification result compared with other state-of-the-art methods in the small training sample size.

Visual attentional-driven deep learning method for flower recognition.

As a typical fine-grained image recognition task, flower category recognition is one of the most popular research topics in the field of computer vision and forestry informatization. Although the image recognition method based on Deep Convolutional Neural Network (DCNNs) has achieved acceptable performance on natural scene image, there are still shortcomings such as lack of training samples, intra-class similarity and low accuracy in flowers category recognition. In this paper, we study deep learning-based flowers' category recognition problem, and propose a novel attention-driven deep learning model to solve it. Specifically, since training the deep learning model usually requires massive training samples, we perform image augmentation for the training sample by using image rotation and cropping. The augmented images and the original image are merged as a training set. Then, inspired by the mechanism of human visual attention, we propose a visual attention-driven deep residual neural network, which is composed of multiple weighted visual attention learning blocks. Each visual attention learning block is composed by a residual connection and an attention connection to enhance the learning ability and discriminating ability of the whole network. Finally, the model is training in the fusion training set and recognize flowers in the testing set. We verify the performance of our new method on public Flowers 17 dataset and it achieves the recognition accuracy of 85.7%.

Phenology-based sample generation for supervised crop type classification

Crop type mapping is relevant to a wide range of food security applications. Supervised classification methods commonly generate these data from satellite image time-series. Yet, their successful implementation is hindered by the lack of training samples. Solutions like transfer learning, development of temporal-spectral signatures of the target classes, re-utilization of existing inventories, or crowdsourcing initiatives are commonly applied to generate samples for thematically coarser classifications. These methods are rarely used for generating crop types samples. In this study, we leverage the phenology information of existing data inventories using Time-Weighted Dynamic Time Warping (TWDTW) to address the problem of automatic crop sample generation in two target areas. Resulting labeled samples are refined using proximity measures obtained from Random Forests (RF). Sentinel-2 time-series are used to obtain phenology information from two study areas. The proposed methodology achieved promising results for classes with a reduced inter-classes similarity such as sugar beets (user’s accuracy, UA, of 98% and producer’s accuracy, PA, of 100%) or grains (UA of 98% and PA of 90%). The crops with a high inter-classes similarity yielded less satisfactory results. Potatoes, for example, obtained a high PA of 95%, but a UA of only 36% because of the spectral-temporal similarity with maize. The methodology works well for areas with balanced crop samples. Yet, it favors prevalent classes in areas with imbalanced crops at the expense of a low accuracy for the minority crops. Despite these shortcomings, the proposed methodology offers a viable option to generate crop samples in regions with few ground labels.

Retrieving the National Main Commodity Maps in Indonesia Based on High-Resolution Remotely Sensed Data Using Cloud Computing Platform

Indonesia has the most favorable climates for agriculture because of its location in the tropical climatic zones. The country has several commodities to support economics growth that are driven by key export commodities—e.g., oil palm, rubber, paddy, cacao, and coffee. Thus, identifying the main commodities in Indonesia using spatially-explicit tools is essential to understand the precise productivity derived from the agricultural sectors. Many previous studies have used predictions developed using binary maps of general crop cover. Here, we present national commodity maps for Indonesia based on remote sensing data using Google Earth Engine. We evaluated a machine learning algorithm—i.e., Random Forest to parameterize how the area in commodity varied in Indonesia. We used various predictors to estimate the productivity of various commodities based on multispectral satellite imageries (36 predictors) at 30-meters spatial resolution. The national commodity map has a relatively high accuracy, with an overall accuracy of about 95% and Kappa coefficient of about 0.90. The results suggest that the oil palm plantation was the highest commodity product that occupied the largest land of Indonesia. However, this study also showed that the land area in rubber, rice paddies, and cacao commodities was underestimated due to its lack of training samples. Improvement in training data collection for each commodity should be done to increase the accuracy of the commodity maps. The commodity data can be viewed online (website can be found in the end of conclusions). This data can further provide significant information related to the agricultural sectors to investigate food provisioning, particularly in Indonesia.

RStaple: A Robust Complementary Learning Method for Real-Time Object Tracking

Object tracking is a challenging research task because of drastic appearance changes of the target and a lack of training samples. Most online learning trackers are hampered by complications, e.g., drifting problem under occlusion, being out of view, or fast motion. In this paper, a real-time object tracking algorithm termed “robust sum of template and pixel-wise learners” (rStaple) is proposed to address those problems. It combines multi-feature correlation filters with a color histogram. Firstly, we extract a combination of specific features from the searching area around the target and then merge feature channels to train a translation correlation filter online. Secondly, the target state is determined by a discriminating mechanism, wherein the model update procedure stops when the target is occluded or out of view, and re-activated when the target re-appears. In addition, by calculating the color histogram score in the searching area, a significant enhancement is adopted for the score map. The target position can be estimated by combining the enhanced color histogram score with the correlation filter response map. Finally, a scale filter is trained for multi-scale detection to obtain the final tracking result. Extensive experimental results on a large benchmark dataset demonstrates that the proposed rStaple is superior to several state-of-the-art algorithms in terms of accuracy and efficiency.

A Dual-Branch Extraction and Classification Method Under Limited Samples of Hyperspectral Images Based on Deep Learning

The convolutional neural network (CNN) has been gradually applied to the hyperspectral images (HSIs) classification, but the lack of training samples caused by the difficulty of HSIs sample marking and ignoring of correlation between spatial and spectral information seriously restrict the HSIs classification accuracy. In an attempt to solve these problems, this paper proposes a dual-branch extraction and classification method under limited samples of hyperspectral images based on deep learning (DBECM). At first, a sample augmentation method based on local and global constraints in this model is designed to augment the limited training samples and balance the number of different class samples. Then spatial-spectral features are simultaneously extracted by the dual-branch spatial-spectral feature extraction method, which improves the utilization of HSIs data information. Finally, the extracted spatial-spectral feature fusion and classification are integrated into a unified network. The experimental results of two typical datasets show that the DBECM proposed in this paper has certain competitive advantages in classification accuracy compared with other public HSIs classification methods, especially in the Indian pines dataset. The parameters of the overall accuracy (OA), average accuracy (AA), and Kappa of the method proposed in this paper are at least 4.7%, 5.7%, and 5% higher than the existing methods.

Design and Implementation of Abnormal Behavior Detection Based on Deep Intelligent Analysis Algorithms in Massive Video Surveillance

Aiming at the high complexity of existing crowd abnormal detection models, the inability of traditional CNN to extract time-related features, and the lack of training samples, an improved spatial-temporal convolution neural network is proposed in this paper. The algorithm firstly uses the aggregation channel feature model to process the surveillance image, and selects the suspected object region with saliency characteristics. Then, the scaled correction and feature extraction are performed on the obtained suspected object region. The corresponding low-level features are obtained and input into the deep network for deep feature learning so as to enhance the representation ability. Finally, the deep feature is input into the least squares SVM classification model to obtain the final abnormal behavior detection result. The embedded chip Hi353I is used as the hardware processor to realize the real-time abnormal behavior detection effect. Our proposed deep intelligent analysis algorithm is used as abnormal Behavior Detector in the board level test. The results show that most of abnormal behaviors can be detected and the alarming message can be timely transmitted in the real-time surveillance.

Classification of EEG signals in epilepsy using a novel integrated TSK fuzzy system

The use of machine learning technology to recognize electrical signals of the brain is becoming increasingly popular. Compared with doctors’ manual judgment, machine learning methods are faster. However, only when its recognition accuracy reaches a high level can it be used in practice. Due to the difference in the data distributions of the training dataset and the test dataset and the lack of training samples, the classification accuracies of general machine learning algorithms are not satisfactory. In fact, among the many machine learning methods used to process epilepsy electroencephalogram (EEG) signals, most are black box methods; however, in medicine, methods with explanatory power are needed. In response to these three challenges, this paper proposes a novel technique based on domain adaptation learning, semi-supervised learning and a fuzzy system. In detail, we use domain adaptation learning to reduce deviation from the data distribution, semi-supervised learning to compensate for the lack of training samples, and the Takagi-Sugen-Kang (TSK) fuzzy system model to improve interpretability. Our experimental results show that the performance of the new method is better than those of most advanced epilepsy classification methods.

Hyperspectral Image Classification Method Based on 2D–3D CNN and Multibranch Feature Fusion

The emergence of a convolutional neural network (CNN) has greatly promoted the development of hyperspectral image (HSI) classification technology. However, the acquisition of HSI is difficult. The lack of training samples is the primary cause of low classification performance. The traditional CNN-based methods mainly use the 2-D CNN for feature extraction, which makes the interband correlations of HSIs underutilized. The 3-D CNN extracts the joint spectral–spatial information representation, but it depends on a more complex model. Also, too deep or too shallow network cannot extract the image features well. To tackle these issues, we propose an HSI classification method based on the 2D–3D CNN and multibranch feature fusion. We first combine 2-D CNN and 3-D CNN to extract image features. Then, by means of the multibranch neural network, three kinds of features from shallow to deep are extracted and fused in the spectral dimension. Finally, the fused features are passed into several fully connected layers and a softmax layer to obtain the classification results. In addition, our network model utilizes the state-of-the-art activation function Mish to further improve the classification performance. Our experimental results, conducted on four widely used HSI datasets, indicate that the proposed method achieves better performance than the existing alternatives.

Deep feature descriptor based hierarchical dense matching for X-ray angiographic images

Backgroud and Objective: X-ray angiography, a powerful technique for blood vessel visualization, is widely used for interventional diagnosis of coronary artery disease because of its fast imaging speed and perspective inspection ability. Matching feature points in angiographic images is a considerably challenging task due to repetitive weak-textured regions.Methods: In this paper, we propose an angiographic image matching method based on the hierarchical dense matching framework, where a novel deep feature descriptor is designed to compute multilevel correlation maps. In particular, the deep feature descriptor is computed by a deep learning model specifically designed and trained for angiographic images, thereby making the correlation maps more distinctive for corresponding feature points in different angiographic images. Moreover, point correspondences are further hierarchically extracted from multilevel correlation maps with the highest similarity response(s), which is relatively robust and accurate. To overcome the problem regarding the lack of training samples, the convolutional neural network (designed for deep feature descriptor) is initially trained on samples from natural images and then fine-tuned on manually annotated angiographic images. Finally, a dense matching completion method, based on the distance between deep feature descriptors, is proposed to generate dense matches between images.Results: The proposed method has been evaluated on the number and accuracy of extracted matches and the performance of subtraction images. Experiments on a variety of angiographic images show promising matching accuracy, compared with state-of-the-art methods.Conclusions: The proposed angiographic image matching method is shown to be accurate and effective for feature matching in angiographic images, and further achieves good performance in image subtraction.

Urban Tree Species Classification Using a WorldView-2/3 and LiDAR Data Fusion Approach and Deep Learning.

Urban areas feature complex and heterogeneous land covers which create challenging issues for tree species classification. The increased availability of high spatial resolution multispectral satellite imagery and LiDAR datasets combined with the recent evolution of deep learning within remote sensing for object detection and scene classification, provide promising opportunities to map individual tree species with greater accuracy and resolution. However, there are knowledge gaps that are related to the contribution of Worldview-3 SWIR bands, very high resolution PAN band and LiDAR data in detailed tree species mapping. Additionally, contemporary deep learning methods are hampered by lack of training samples and difficulties of preparing training data. The objective of this study was to examine the potential of a novel deep learning method, Dense Convolutional Network (DenseNet), to identify dominant individual tree species in a complex urban environment within a fused image of WorldView-2 VNIR, Worldview-3 SWIR and LiDAR datasets. DenseNet results were compared against two popular machine classifiers in remote sensing image analysis, Random Forest (RF) and Support Vector Machine (SVM). Our results demonstrated that: (1) utilizing a data fusion approach beginning with VNIR and adding SWIR, LiDAR, and panchromatic (PAN) bands increased the overall accuracy of the DenseNet classifier from 75.9% to 76.8%, 81.1% and 82.6%, respectively. (2) DenseNet significantly outperformed RF and SVM for the classification of eight dominant tree species with an overall accuracy of 82.6%, compared to 51.8% and 52% for SVM and RF classifiers, respectively. (3) DenseNet maintained superior performance over RF and SVM classifiers under restricted training sample quantities which is a major limiting factor for deep learning techniques. Overall, the study reveals that DenseNet is more effective for urban tree species classification as it outperforms the popular RF and SVM techniques when working with highly complex image scenes regardless of training sample size.

Active Learning for Uneven Noisy Labeled Data in Mention-Level Relation Extraction

Mention-level relation extraction (mRE) plays an important role in extracting relational information from short texts such as those exchanged in a social network. Deep learning (DL) has made remarkable achievements; the main problem encountered with DL in mRE is a lack of training samples. In this paper, we present a design for a quick sample-marking method. First, we construct an uneven noisy labeled data (UNLD) set using a pattern matching algorithm, and then a relabeling framework is put forward for modifying the UNLD. With regard to the accuracy, the recall rates of categories with sufficient samples increased from 0.4 to nearly 1 using the relabeling framework. We have released our code and other resources for further research (https://github.com/curtainsky/UNLD).

Hyperspectral Image Classification Based on Deep Deconvolution Network With Skip Architecture

Convolution neural network (CNN) utilizes alternating convolutional and pooling layers to learn representative spatial information when the training samples are sufficient. However, for pixelwise classification of hyperspectral image, some important information is neglected by CNN, such as the erased information by the pooling operation and the appearance information from lower layers. Moreover, the lack of training samples is a common situation in remote sensing area, which afflicts CNN with overfitting problem. To address the aforementioned issues, this paper designs an end-to-end deconvolution network with skip architecture to learn the spectral–spatial features. The proposed network starts with two branches, i.e., the spatial branch and spectral branch. In the spatial branch, a band selection layer is designed to reduce parameters and remit the overfitting problem, unpooling and deconvolution operations are utilized to recover the erased information of the pooling layers and learn pixelwise spatial representation hierarchically, and the skip architecture is constructed for merging the deep semantic information with the shallow appearance information. In the spectral branch, a contextual deep network is employed for learning deep spectral features. Experimental results on three benchmark data sets reveal the competitive performance of the proposed approach over several related methods.

An Improving EFA for Clutter Suppression by Using the Persymmetric Covariance Matrix Estimation

The extended factored approach (EFA) is believed to be one of the most efficient and practical space–time adaptive processing (STAP) algorithms for clutter suppression in an airborne radar system. However, it cannot effectively work in the airborne radar system with large antenna array for the huge computational cost and the lack of training sample. To solve these problems, a bi-iterative algorithm based on the persymmetric covariance matrix estimation is proposed in this paper. Firstly, the clutter covariance matrix is estimated by using the original data, the constructed spatial transformed data, the constructed temporal transformed data and the constructed spatial–temporal transformed data. Secondly, the spatial weight vector in EFA is decomposed as the Kronecker products of two short weight vectors. Finally, the bi-iterative algorithm is exploited to obtain the desired weight vectors. Thus, the improving EFA with small training sample demanding is realized. Experimental results demonstrate the effectiveness of the proposed method under small training sample support.

SAR Unlabeled Target Recognition Based on Updating CNN With Assistant Decision[-.2pc

Compared with optical images, the number of synthetic aperture radar (SAR) images is limited and growing slowly over time, and most of time the additional SAR images are unlabeled, which restricts the development of target recognition on SAR images. In this letter, the update learning method for SAR unlabeled target recognition based on the convolutional neural network (CNN) with assistant decision is proposed. The seed image set, which contains a small amount of existing labeled SAR images, is used for pretraining a CNN model. The unlabeled samples are fed to both CNN and the assistant classifier, and then, the matrices of the unlabeled samples based on recognition probability will be obtained, respectively. The unlabeled samples with high confidence by decision will be fed to CNN again to fine-tune the CNN model. The proposed method can make full use of the unlabeled samples to update the recognition model and to improve the recognition performance. Experiments on MSTAR data set show that, with the increase of unlabeled samples, the proposed method can improve the recognition performance and outperform the methods of CNN self-update learning and support vector machine update learning with the principal component analysis and the nonnegative matrix factorization. The experiments’ results demonstrate that the proposed method can effectively solve the problems caused by the lack of training samples.

Palmprint Identification Using an Ensemble of Sparse Representations

Among various palmprint identification methods proposed in the literature, sparse representation for classification (SRC) is very attractive offering high accuracy. Although SRC has good discriminative ability, its performance strongly depends on the quality of the training data. In particular, SRC suffers from two major problems: lack of training samples per class and large intra-class variations. In fact, palmprint images not only contain identity information but they also have other information, such as illumination and geometrical distortions due to the unconstrained conditions and the movement of the hand. In this case, the sparse representation assumption may not hold well in the original space since samples from different classes may be considered from the same class. This paper aims to enhance palmprint identification performance through SRC by proposing a simple yet efficient method based on an ensemble of sparse representations through an ensemble of discriminative dictionaries satisfying SRC assumption. The ensemble learning has the advantage to reduce the sensitivity due to the limited size of the training data and is performed based on random subspace sampling over 2D-PCA space while keeping the image inherent structure and information. In order to obtain discriminative dictionaries satisfying SRC assumption, a new space is learned by minimizing and maximizing the intra-class and inter-class variations using 2D-LDA. Extensive experiments are conducted on two publicly available palmprint data sets: multispectral and PolyU. Obtained results showed very promising results compared with both state-of-the-art holistic and coding methods. Besides these findings, we provide an empirical analysis of the parameters involved in the proposed technique to guide the neophyte.

An improved adaptation algorithm for signer-independent sign language recognition

Sign language recognition is a technology that can present sign language in a understandable form, in order to achieve barrier-free communication between deaf and normal. To solve the differences in sign language data issues and the lack of training samples of manpower caused by low recognition of non-specific language, presents MLLR\MAP adaptive progressive non-specific integrated manpower language recognition framework. This approach optimises the division MLLR regression class to provide more accurate initial MAP model, which give full play to the rapidity and the MAP MLLR progressive. Then introduced MCE model parameter estimation algorithm to compensate for the limitations of the model parameters adaptive method to further reduce the system error rate and accelerate the recognition speed. Meanwhile, for the MCE algorithm computationally intensive problems proposed improvements. Experimental results show that the adaptive sign language data required for this algorithm is less than traditional MLLR and MAP methods, while improved average recognition rate by 15.6%.