Generated Training Samples Research Articles

TAILORED FEATURES FOR SEMANTIC SEGMENTATION WITH A DGCNN USING FREE TRAINING SAMPLES OF A COLORED AIRBORNE POINT CLOUD

Abstract. Automation of 3D LiDAR point cloud processing is expected to increase the production rate of many applications including automatic map generation. Fast development on high-end hardware has boosted the expansion of deep learning research for 3D classification and segmentation. However, deep learning requires large amount of high quality training samples. The generation of training samples for accurate classification results, especially for airborne point cloud data, is still problematic. Moreover, which customized features should be used best for segmenting airborne point cloud data is still unclear. This paper proposes semi-automatic point cloud labelling and examines the potential of combining different tailor-made features for pointwise semantic segmentation of an airborne point cloud. We implement a Dynamic Graph CNN (DGCNN) approach to classify airborne point cloud data into four land cover classes: bare-land, trees, buildings and roads. The DGCNN architecture is chosen as this network relates two approaches, PointNet and graph CNNs, to exploit the geometric relationships between points. For experiments, we train an airborne point cloud and co-aligned orthophoto of the Surabaya city area of Indonesia to DGCNN using three different tailor-made feature combinations: points with RGB (Red, Green, Blue) color, points with original LiDAR features (Intensity, Return number, Number of returns) so-called IRN, and points with two spectral colors and Intensity (Red, Green, Intensity) so-called RGI. The overall accuracy of the testing area indicates that using RGB information gives the best segmentation results of 81.05% while IRN and RGI gives accuracy values of 76.13%, and 79.81%, respectively.

Immune-related genes have prognostic significance in head and neck squamous cell carcinoma

AimsHead and neck squamous cell carcinoma (HNSCC) is an highly aggressive tumor with heterogeneous prognosis. We here report that immune-related genes (IRGs) could effectively distinguish prognostically different HNSCC patients. Materials and methodsMRNA levels of 1333 IRGs that from ImmPort database in HNSCC samples were acquired from the Cancer Genome Atlas (TCGA). H2o, a machine learning-based R package, was used for screening the top most representative genes from the IRGs. Univariate Cox-regression analysis was performed to identify prognostically-related genes based on the randomly generated training samples from TCGA set. LASSO Cox-regression analysis was applied for the construction of prognostic model for HNSCC. A total of six IRGs were finally retained for their prognostic significance and used for LASSO Cox-regression analysis. Key findingsSamples from exclusive training and testing set that randomly generated from TCGA, and another independent validation set from the Gene Expression Omnibus (GEO) were divided into high- and low-risk groups according to the prognostic model. HNSCC samples within high-risk groups have significantly inferior overall survival (OS) compared with those within low-risk groups. Differences in genomic mutation landscape and tumor infiltration immune cells also exist between the two sample groups. What's more, risk score was proved to be an independent prognostic factor for HNSCC by stratification analysis. SignificanceIRGs are pivotal HNSCC prognostic signatures and should be helpful for its clinical decision-making.

METHOD FOR SELECTING A CLASSIFIED TRAINING SAMPLE FOR RADAR’S ELEVATION CHANNELS

Context. Under conditions of simultaneous exposure to active and passive interference, the passive component decorrelates the active component of the combined interference in the receiving channels of the surveillance radar, which significantly reduces the effectiveness of its suppression. Therefore, in order to adapt the weight coefficients of the spatial filter during spatio-temporal processing of radar signals, it is necessary, by one method or another, to form a classified training sample generated only by active noise interference.Objective. The goal is to optimize the choice of a method for generating a classified training sample for various elevation channels of the radar field of view, taking into account the predicted distribution of passive interference in range.Method. An analytical assessment of the distribution of the most powerful Cb type cumulonimbus cloud systems, followed by mathematical modeling of the effectiveness of the active noise suppression using various training samples generation methods.Results. Based on the performed analytical calculations, it was shown that in the upper part of the field of view of surveillance radar in the sector of elevation angles from 6º to 30º, the upper boundary of Cb type clouds does not exceed 75 km, which guarantees the absence of passive interference at the end of the radar range. This makes it possible to form a training sample at a time interval located before the radiation of the next probe pulse. In the lower part of the field of view in the sector of angles 0º to 6º, the upper boundary of the clouds is located at a distance of 75–270 km, which does not allow the formation of a training sample at the end of the radar range. It is proposed to use the correlation method, in which the time interval for the formation of the training sample is determined by the “on the fly” analysis of the magnitude of the inter-channel correlation coefficient. The choice of the interval with the highest value of inter-channel correlation during the repetition period allows us to form a training sample with the least influence of passive interference on the process of suppressing the active component of the combined interference. As a result of modeling, it was found that the use of a training sample significantly reduces the duration of the transient process when forming the weight coefficients of the spatial filter and, accordingly, improves the quality of the suppression of active interference.Conclusions. The scientific novelty of the work lies in the fact that for the first time the problem of individual selection of a classified training sample for the surveillance radar elevation channels has been solved. Practical significance consists in the development of a methodology for analyzing the location of the most powerful cloud systems in range in the radar field of view with varying degrees of influence of passive interference in elevation channels in a real signal-noise environment. To the practical novelty should also include the results of simulation modeling, which confirm the need to choose a method for generating a classified training sample in different elevation channels of the radar viewing area should also be attributed.

An ultrafast and high accuracy calculation method for gas radiation characteristics using artificial neural network

In this work, a new approach to efficient gas radiation characteristics calculating is proposed to satisfy the demand for high accuracy and efficient calculation in many applications. This approach establishes a mapping relationship between gas condition parameters and radiation characteristics using a back-propagation neural network (BPNN). The line by line (LBL) model is utilized for the generation of training samples in the BPNN model. The values of pressure, temperature and component concentration are taken as input, and absorption coefficient values are taken as output. A case study of CO2 transmittance at 2250–2350 cm− 1 band is presented. The comparison and analysis of the results indicated that the BPNN model has a high accuracy of LBL fitting and is insensitive to the input. Although the training time of BPNN is long, once the training is completed, the computational efficiency is very high. Compared to the look-up table method or other accelerated methods using parameter pre-calculation, the BPNN method occupies much less storage space. It can replace the LBL model to a certain extent when dealing with the needs of high precision and high-speed computing.

Isocitrate dehydrogenase (IDH) status prediction in histopathology images of gliomas using deep learning

Mutations in isocitrate dehydrogenase genes IDH1 and IDH2 are frequently found in diffuse and anaplastic astrocytic and oligodendroglial tumours as well as in secondary glioblastomas. As IDH is a very important prognostic, diagnostic and therapeutic biomarker for glioma, it is of paramount importance to determine its mutational status. The haematoxylin and eosin (H&E) staining is a valuable tool in precision oncology as it guides histopathology-based diagnosis and proceeding patient’s treatment. However, H&E staining alone does not determine the IDH mutational status of a tumour. Deep learning methods applied to MRI data have been demonstrated to be a useful tool in IDH status prediction, however the effectiveness of deep learning on H&E slides in the clinical setting has not been investigated so far. Furthermore, the performance of deep learning methods in medical imaging has been practically limited by small sample sizes currently available. Here we propose a data augmentation method based on the Generative Adversarial Networks (GAN) deep learning methodology, to improve the prediction performance of IDH mutational status using H&E slides. The H&E slides were acquired from 266 grade II-IV glioma patients from a mixture of public and private databases, including 130 IDH-wildtype and 136 IDH-mutant patients. A baseline deep learning model without data augmentation achieved an accuracy of 0.794 (AUC = 0.920). With GAN-based data augmentation, the accuracy of the IDH mutational status prediction was improved to 0.853 (AUC = 0.927) when the 3,000 GAN generated training samples were added to the original training set (24,000 samples). By integrating also patients’ age into the model, the accuracy improved further to 0.882 (AUC = 0.931). Our findings show that deep learning methodology, enhanced by GAN data augmentation, can support physicians in gliomas’ IDH status prediction.

Harnessing GANs for Zero-Shot Learning of New Classes in Visual Speech Recognition

Visual Speech Recognition (VSR) is the process of recognizing or interpreting speech by watching the lip movements of the speaker. Recent machine learning based approaches model VSR as a classification problem; however, the scarcity of training data leads to error-prone systems with very low accuracies in predicting unseen classes. To solve this problem, we present a novel approach to zero-shot learning by generating new classes using Generative Adversarial Networks (GANs), and show how the addition of unseen class samples increases the accuracy of a VSR system by a significant margin of 27% and allows it to handle speaker-independent out-of-vocabulary phrases. We also show that our models are language agnostic and therefore capable of seamlessly generating, using English training data, videos for a new language (Hindi). To the best of our knowledge, this is the first work to show empirical evidence of the use of GANs for generating training samples of unseen classes in the domain of VSR, hence facilitating zero-shot learning. We make the added videos for new classes publicly available along with our code1.

Using Training Samples Retrieved from a Topographic Map and Unsupervised Segmentation for the Classification of Airborne Laser Scanning Data

The labeling of point clouds is the fundamental task in airborne laser scanning (ALS) point clouds processing. Many supervised methods have been proposed for the point clouds classification work. Training samples play an important role in the supervised classification. Most of the training samples are generated by manual labeling, which is time-consuming. To reduce the cost of manual annotating for ALS data, we propose a framework that automatically generates training samples using a two-dimensional (2D) topographic map and an unsupervised segmentation step. In this approach, input point clouds, at first, are separated into the ground part and the non-ground part by a DEM filter. Then, a point-in-polygon operation using polygon maps derived from a 2D topographic map is used to generate initial training samples. The unsupervised segmentation method is applied to reduce the noise and improve the accuracy of the point-in-polygon training samples. Finally, the super point graph is used for the training and testing procedure. A comparison with the point-based deep neural network Pointnet++ (average F1 score 59.4%) shows that the segmentation based strategy improves the performance of our initial training samples (average F1 score 65.6%). After adding the intensity value in unsupervised segmentation, our automatically generated training samples have competitive results with an average F1 score of 74.8% for ALS data classification while using the ground truth training samples the average F1 score is 75.1%. The result shows that our framework is feasible to automatically generate and improve the training samples with low time and labour costs.

Teaching machines to write like humans using L-attributed grammar

Reading and writing are easy for humans. The automatic reading of handwritten characters has been studied for several decades. Machine learning algorithms for reading tasks often require a huge amount of data to perform with similar accuracy to humans, yet it is also difficult to gain sufficient meaningful data. Automatic writing tasks have not been studied as extensively. In this paper, we teach machines to write like teaching a child by telling the machine the method for writing each character using L-attributed grammar. With the aid of the proposed TMTW (Teaching Machines To Write) interacting system, a human as a teacher only needs to provide the writing sequence of parts and control lines. The proposed system automatically perceives the relationships between control lines and parts, and constructs the grammars. Top-down derivation and the stroke generation method are applied to generate varying characters based on the learned grammars. For as long as a machine can write, it can be applied in robot control or training sample generation for automatic reading tasks. The MNIST and CASIA datasets are used to demonstrate the effectiveness of the proposed system on different languages. The machine written samples are used to train a network, which is evaluated on the MNIST test set. A test error rate of 1.23% is achieved using only approximately 20 grammars on average for each digit. Using the generated and handwritten samples together as a training set can reduce the test error rate to 0.61%. Similar experiments are conducted using the CASIA data set, and the results demonstrated that the proposed method is effective in generating characters with a complex structure. The source codes and grammars used in this paper have been made publicly available in https://github.com/step123456789/TMTW.

New Scheme for Impervious Surface Area Mapping From SAR Images With Auxiliary User-Generated Content

This article presents a new scheme to extract impervious surface area from synthetic-aperture radar (SAR) images exploiting auxiliary user-generated content (UGC). The presented scheme includes the automatic generation of training samples based on the combination of UGC and SAR data, and SAR data preprocessing, leading to impervious surface area classification through a clustering-based one-class support vector machine approach. Two areas—namely, the cities of Beijing and Taipei, have been analyzed using the Sentinel-1 SAR data to test and validate the proposed methodology. Experimental results show that the presented scheme improves the automatic selection of impervious surface training samples. Moreover, this scheme achieves a comparable classification performance to traditional methods without requiring time-consuming training point manual extraction. Results in this study will help to promote the application of UGC for urban remote sensing data interpretation.

A New Nonparametric Algorithm for Preprocessing Stochastic Data with Uncertainty

The article deals with the problem of modeling stochastic processes under uncertainty. The peculiarity of the processes under consideration is that the researcher does not have information about the mathematical structure of the object; the object is represented as a black box. The article proposes to use a nonparametric modeling algorithm based on a nonparametric estimate of the regression function on observations. To improve the accuracy of modeling, it is proposed to use an algorithm for generating training samples. The algorithm differs from the previous modification by the definition of essential variables. The results of computational experiments have shown the effectiveness of the proposed algorithms.

Синтез изображений интерьеров для обучения нейросетей

The paper proposes a number of methods that can be used to synthesize images of interiors in order to train artificial intelligence. The proposed methods solve the problem of generating training samples in a complex, starting from automatic generation of 3D content and ending with rendering directly. One of the main goals of the develioed system is to provide sufficient performance when generating sets of photo-realistic images of interiors via using GPUs.

Ship Detection Using Deep Convolutional Neural Networks for PolSAR Images

Ship detection plays an important role in many remote sensing applications. However, the performance of the PolSAR ship detection may be degraded by the complicated scattering mechanism, multi-scale size of targets, and random speckle noise, etc. In this paper, we propose a ship detection method for PolSAR images based on modified faster region-based convolutional neural network (Faster R-CNN). The main improvements include proposal generation by adopting multi-level features produced by the convolution layers, which fits ships with different sizes, and the addition of a Deep Convolutional Neural Network (DCNN)-based classifier for training sample generation and coast mitigation. The proposed method has been validated by four measured datasets of NASA/JPL airborne synthetic aperture radar (AIRSAR) and uninhabited aerial vehicle synthetic aperture radar (UAVSAR). Performance comparison with the modified constant false alarm rate (CFAR) detector and the Faster R-CNN has demonstrated that the proposed method can improve the detection probability while reducing the false alarm rate and missed detections.

Accelerated Multifidelity Emulator Modeling for Probabilistic Rotor Response Study

Abstract This paper presents an efficient methodology to build a modal solution emulator for the probabilistic study of geometrically mistuned bladed rotors by using the newly developed localized-Galerkin multifidelity (LGMF) modeling and eigensolution reanalysis (ER) with the symmetric successive matrix inversion (SSMI) methods. The key idea of the mistuned blade emulator is to establish a reduced functional relationship between the stochastic geometric variations and the disturbed modal responses. The prediction accuracy of an emulator generally depends on how many training samples of modal solutions are available and how well the potential modal switching due to stochastic mistuning is captured. To reduce the computational costs of generating training samples without sacrificing accuracy, this paper introduces the collaborative framework of the new approaches of multifidelity (MF) modeling and ER. The proposed framework is demonstrated for its computational benefits with several numerical examples including the point-cloud scanned mistuned blade problem.

A graded proportion method of training sample selection for updating conventional soil maps

Selection of training samples is a vital step in updating conventional soil maps when utilizing data mining models. Quality of training samples significantly affects the mapping results and accuracies of the updated soil maps. The area-weighted proportion method was a common method for generating training samples. However, this method usually assigns too small weight to those soil types of small areas and large weight to those of large areas in sample size allocation, which causes the unreasonable proportions of sample numbers for soil types and thereby biases the representation of soil-environmental relationships for those soil types. Meanwhile, random selection of training samples from a soil type may generate some ‘noise’ samples located in the transition areas between soil types. These two aspects in training sample selection could probably reduce the accuracy of the updated soil maps. In this study, a new method was developed to select training samples based on soil type grading according to their area coverages. The method consists of two steps. The first step is to determine the numbers of training samples for each soil type based on soil type grading so as to maintain the reasonable proportion in sample numbers among soil types with different area coverages. The second step is to select typical (representative) samples for each soil type from conventional soil map, to avoid generation of ‘noise samples’. To evaluate the proposed method, the method was compared with three other training sample selection methods with four training sample sizes. Each method was ran for 100 times to generate training sample datasets with each sample size to evaluate their effectiveness and stability. Random forest was employed to generate updated soil maps in a small watershed in Raffelson, Wisconsin (USA). The validation results showed that the graded proportion method effectively solved the imbalanced issue of training samples among soil types with area coverages in big differences caused by the area-weighted proportion strategy. Thus training samples generated using the proposed method usually obtained more accurate and reasonable mapping results than those using the area-weighted proportion strategy. Furthermore, the performance of the proposed method was more stable than that of the area-weighted proportion strategy with the training sample size increasing. It is concluded that the proposed method is an effective training sample selection method for data mining model to update conventional soil maps.

Crop Mapping Based on Historical Samples and New Training Samples Generation in Heilongjiang Province, China

Accurate, year-by-year crop distribution information is a key element in agricultural production regulation and global change governance. However, due to the high sampling costs and insufficient use of historical samples, a supervised classifying method for sampling every year is unsustainable for mapping crop types over time. Therefore, this paper proposes a method for the generation and screening of new samples for 2018 based on historical crop samples, and then it builds a crop mapping model for that current season. Pixels with the same crop type in the historical year (2013–2017) were extracted as potential samples, and their spectral features and spatial information in the current year (2018) were used to generate new samples based on clustering screening. The research result shows that when the clustering number is different, the number and structure of new generated sample also changes. The sample structure generated in Luobei County was not balanced, with the ‘other crop’ representing less than 3.97%, but the structure of southwest Hulin City was more balanced. Based on the newly generated samples and the ground reference data of classified year, the classification models were constructed. The average classification accuracies of Luobei County in 2018 based on new generated samples and field samples were 69.35% and 77.59%, respectively, while those of southwest Hulin City were 80.44% and 82.94%, respectively. Combined with historical samples and the spectral information of the current year, this study proposes a method to generate new samples. It can overcome the problem of crop samples only being collected in the field due to the difficulty of visual interpretation, effectively improve the use of historical data, and also provide a new idea for sustainable crop mapping in many regions lacking seasonal field samples.

Individual Rubber Tree Segmentation Based on Ground-Based LiDAR Data and Faster R-CNN of Deep Learning

Rubber trees in southern China are often impacted by natural disturbances that can result in a tilted tree body. Accurate crown segmentation for individual rubber trees from scanned point clouds is an essential prerequisite for accurate tree parameter retrieval. In this paper, three plots of different rubber tree clones, PR107, CATAS 7-20-59, and CATAS 8-7-9, were taken as the study subjects. Through data collection using ground-based mobile light detection and ranging (LiDAR), a voxelisation method based on the scanned tree trunk data was proposed, and deep images (i.e., images normally used for deep learning) were generated through frontal and lateral projection transform of point clouds in each voxel with a length of 8 m and a width of 3 m. These images provided the training and testing samples for the faster region-based convolutional neural network (Faster R-CNN) of deep learning. Consequently, the Faster R-CNN combined with the generated training samples comprising 802 deep images with pre-marked trunk locations was trained to automatically recognize the trunk locations in the testing samples, which comprised 359 deep images. Finally, the point clouds for the lower parts of each trunk were extracted through back-projection transform from the recognized trunk locations in the testing samples and used as the seed points for the region’s growing algorithm to accomplish individual rubber tree crown segmentation. Compared with the visual inspection results, the recognition rate of our method reached 100% for the deep images of the testing samples when the images contained one or two trunks or the trunk information was slightly occluded by leaves. For the complicated cases, i.e., multiple trunks or overlapping trunks in one deep image or a trunk appearing in two adjacent deep images, the recognition accuracy of our method was greater than 90%. Our work represents a new method that combines a deep learning framework with point cloud processing for individual rubber tree crown segmentation based on ground-based mobile LiDAR scanned data.

Abdominal artery segmentation method from CT volumes using fully convolutional neural network.

PURPOSE: The purpose of this paper is to present a fully automated abdominal artery segmentation method from a CT volume. Three-dimensional (3D) blood vessel structure information is important for diagnosis and treatment. Information about blood vessels (including arteries) can be used in patient-specific surgical planning and intra-operative navigation. Since blood vessels have large inter-patient variations in branching patterns and positions, a patient-specific blood vessel segmentation method is necessary. Even though deep learning-based segmentation methods provide good segmentation accuracy among large organs, small organs such as blood vessels are not well segmented. We propose a deep learning-based abdominal artery segmentation method from a CT volume. Because the artery is one of small organs that is difficult to segment, we introduced an original training sample generation method and a three-plane segmentation approach to improve segmentation accuracy. METHOD: Our proposed method segments abdominal arteries from an abdominal CT volume with a fully convolutional network (FCN). To segment small arteries, we employ a 2D patch-based segmentation method and an area imbalance reduced training patch generation (AIRTPG) method. AIRTPG adjusts patch number imbalances between patches with artery regions and patches without them. These methods improved the segmentation accuracies of small artery regions. Furthermore, we introduced a three-plane segmentation approach to obtain clear 3D segmentation results from 2D patch-based processes. In the three-plane approach, we performed three segmentation processes using patches generated on axial, coronal, and sagittal planes and combined the results to generate a 3D segmentation result. RESULTS: The evaluation results of the proposed method using 20 cases of abdominal CT volumes show that the averaged F-measure, precision, and recall rates were 87.1%, 85.8%, and 88.4%, respectively. This result outperformed our previous automated FCN-based segmentation method. Our method offers competitive performance compared to the previous blood vessel segmentation methods from 3D volumes. CONCLUSIONS: We developed an abdominal artery segmentation method using FCN. The 2D patch-based and AIRTPG methods effectively segmented the artery regions. In addition, the three-plane approach generated good 3D segmentation results.

Robust Visual Tracking based on Adversarial Unlabeled Instance Generation with Label Smoothing Loss Regularization

Recent studies have shown that deep neural networks have pushed visual tracking accuracy to new heights, but finding more robust long-term tracking is still challenging because of the dynamic foreground and background changes. This phenomenon affects the overall performance via online training sample generation. The dense sampling strategy has been widely used for its convenience, the appearance variation is severely limited by its highly spatial overlapping mechanism. The sample candidate evaluation with a classification score metric is not always reliable throughout the entire process, therefore, tracking failure is inevitable. As an effective solution, this paper proposes a novel sample-level generative adversarial network (GAN) to enrich the training data by generating massive amounts of sample-level GAN samples. These samples are not only similar to the real-life scenarios, but also could carry more diversity of deformation and motion blur to a certain degree. For occlusion invariance, a feature-level GAN is incorporated to generate more challenging feature-level GAN data by creating random occlusion masks in deep feature space. To facilitate the online learning process, a label smoothing loss regularization is introduced to achieve model regularization and over-fitting reduction by integrating the unlabeled GAN-generated training data with the realistically labeled ones. In addition, a re-detection correlation filter conservatively trained with reliable training data is employed to integrate a classification score metric to perform reliable model updates and avoid heavy degradation. Furthermore, we also carry out the re-detection correlation filter on the candidate region proposals to handle the tracking failures. The proposed tracker has shown superior performance in comparison to the other state-of-the-art tracking approaches on the OTB-2013, OTB-100, UAV123, UAV20L, and VOT2016 benchmark datasets.

An Improved Perceptual Hash Algorithm Based on U-Net for the Authentication of High-Resolution Remote Sensing Image

Data security technology is of great significance for the effective use of high-resolution remote sensing (HRRS) images in GIS field. Integrity authentication technology is an important technology to ensure the security of HRRS images. Traditional authentication technologies perform binary level authentication of the data and cannot meet the authentication requirements for HRRS images, while perceptual hashing can achieve perceptual content-based authentication. Compared with traditional algorithms, the existing edge-feature-based perceptual hash algorithms have already achieved high tampering authentication accuracy for the authentication of HRRS images. However, because of the traditional feature extraction methods they adopt, they lack autonomous learning ability, and their robustness still exists and needs to be improved. In this paper, we propose an improved perceptual hash scheme based on deep learning (DL) for the authentication of HRRS images. The proposed method consists of a modified U-net model to extract robust feature and a principal component analysis (PCA)-based encoder to generate perceptual hash values for HRRS images. In the training stage, a training sample generation method combining artificial processing and Canny operator is proposed to generate robust edge features samples. Moreover, to improve the performance of the network, exponential linear unit (ELU) and batch normalization (BN) are applied to extract more robust and accurate edge feature. The experiments have shown that the proposed algorithm has almost 100% robustness to format conversion between TIFF and BMP, LSB watermark embedding and lossless compression. Compared with the existing algorithms, the robustness of the proposed algorithm to lossy compression has been improved, with an average increase of 10%. What is more, the algorithm has good sensitivity to detect local subtle tampering to meet the high-accuracy requirements of authentication for HRRS images.

Learning a Compact Vein Discrimination Model With GANerated Samples

Despite the great success achieved by convolutional neural networks (CNNs) in various image understanding tasks, it is still difficult for CNNs to be applied to vein recognition tasks due to the problems of insufficient training datasets, intra-class variations, and inter-class similarities. Besides, due to the essential requirement on the storage of millions of parameters for CNN, it is challenging to use a CNN for designing a vein-based embedded person identification system. In this paper, these two problems are addressed by learning a discriminative and compact vein recognition model. For the first problem, a hierarchical generative adversarial network (HGAN) consisting of a constrained CNN and a CycleGAN is proposed for data augmentation. Two similarity losses are defined for estimating the self-similarity and inter-class dissimilarity, and a CycleGAN model is properly trained with these two losses for better task-specific training sample generation. After obtaining a baseline vein recognition model fine-tuned on the augmented datasets, the existence of parameter redundancy in the over-parameterized network motivates the proposal of model compression by way of filter pruning and low rank approximation, thus making the compressed model more suitable for deployment on embedded systems. Through the vein recognition experiments with two different datasets and an additional palmprint recognition experiment, the proposed algorithms are shown to yield a highly compact model while keeping the accuracy acceptable for application.