Balanced Training Samples Research Articles

Balanced parametric body prior for implicit clothed human reconstruction from a monocular RGB

AbstractThe authors study the problem of reconstructing detailed 3D human surfaces in various poses and clothing from images. The parametric human body allows accurate 3D clothed human reconstruction. However, the offset of large and loose clothing from the inferred parametric body mesh confines the generalisation of the existing parametric body‐based methods. A distinctive method that simultaneously generalises well to unseen poses and unseen clothing is proposed. The authors first discover the unbalanced nature of existing implicit function‐based methods. To address this issue, the authors propose to synthesise the balanced training samples with a new dependency coefficient in training. The dependency coefficient can tell the network whether the prior from the parametric body model is reliable. The authors then design a novel positional embedding‐based attenuation strategy to incorporate the dependency coefficient into the implicit function (IF) network. Comprehensive experiments are conducted on the CAPE dataset to study the effectiveness of the authors’ approach. The proposed method significantly surpasses state‐of‐the‐art approaches and generalises well on unseen poses and clothing. As an illustrative example, the proposed method improves the Chamfer Distance Error and Normal Error by 38.2% and 57.6%.

Style-KD: Class-imbalanced medical image classification via style knowledge distillation

Problem:In the medical domain, obtaining training images on a large scale is difficult due to privacy and cost issues, and as further disease incidence varies widely, class imbalance has been a critical problem. Importantly, securing samples for rare diseases will always be difficult, but a diagnosis model should be able to predict the diseased samples accurately. When a model is trained on imbalanced data, it is expected that (1) the model only exhibits biased results for the majority classes, and (2) the recognition performance for minority classes will decline rapidly, resulting in low specificity. Objective:We aim to improve the prediction performance of such few sample classes using our auxiliary data from an efficient generation model and style knowledge distillation. In this paper, we introduce a style knowledge distillation method to address the class imbalance problem. Methods:The proposed method generates auxiliary images in which the style of a training sample is transferred. We train our model with balanced training samples by distilling the styles of auxiliary images. Conclusion:We demonstrate that it considerably increases the accuracy of classes with few samples, notably without adversely impacting classes with large samples, using the APTOS2019, and ISIC2018. This approach is straightforward but very effective for medical image datasets with skewed class distributions.

Dual-Branch Fourier-Mixing Transformer Network for Hyperspectral Target Detection

In recent years, transformers have shown great potential in hyperspectral image processing and have also been gradually applied in hyperspectral target detection (HTD). Nonetheless, applying a typical transformer to HTD remains challenging. The heavy computation burden of the multi-head self-attention (MSA) in transformers limits its efficient HTD, while the limited ability to extract local spectral features can reduce the discrimination of the learned spectral features. To further explore the potential of transformers for HTD, for balance of representation ability and computational efficiency, we propose a dual-branch Fourier-mixing transformer network for hyperspectral target detection (DBFTTD). First, this work explores a dual-branch Fourier-mixing transformer network. The transformer-style network replaces the MSA sublayer in the transformer with a Fourier-mixing sublayer, which shows advantages in improving computational efficiency and learning valuable spectral information effectively for HTD. Second, this work proposes learnable filter ensembles in the Fourier domain that are inspired by ensemble learning to improve detection performance. Third, a simple but efficient dropout strategy is proposed for data augmentation. Sufficient and balanced training samples are constructed for training the dual-branch network, and training samples for balanced learning can further improve detection performance. Experiments on four data sets indicate that our proposed detector is superior to the state-of-the-art detectors.

A novel hierarchical training architecture for Siamese Neural Network based fault diagnosis method under small sample

Although current deep learning-based fault diagnosis methods have made great progress, the accuracy of these models is usually attained based on many balanced training samples, which is quite difficult to meet in real industrial scenarios. In this paper, a Siamese Neural Network (SNN) based fault diagnosis method under small sample conditions is proposed and a novel hierarchical training architecture is designed to train model efficiently. Through pre-training and secondary training of the feature extractor, the SNN training stagnation problem caused by input data is addressed, which fully exploits the deep features of small sample data. Then, a classifier is added to construct the final fault diagnosis network, which improves the diagnosis speed under large test samples, and realizes accurate and efficient fault diagnosis under small sample conditions. Finally, experiments using real rotating machinery datasets show the feasibility and accuracy of the proposed method under small sample.

Memory-Augmented Point Cloud Registration Network for Bucket Pose Estimation of the Intelligent Mining Excavator

Bucket pose estimation is a key technology for the intelligent mining excavator to realize automatic excavation and loading. The existing estimation methods mainly include two categories. The sensor measurement methods install sensors to detect the bucket pose, which suffer from large cumulative error and short service life caused by strong vibrations. The traditional registration methods estimate the bucket pose by registering its point clouds with the iterative closest point method, which easily fall into a local optimum. Recently, point cloud registration networks have been studied for object pose estimation based on balanced training samples. However, since the bucket speed changes obviously during excavation, the number of bucket samples at different poses is imbalanced. As a result, point cloud registration networks will easily forget the bucket features at some poses which have small samples, also called nondominant pose features, during learning. This will reduce the accuracy of bucket pose estimation. To handle the forgetting problem caused by the sample imbalance, we propose a novel memory-augmented registration network (MARNet) for bucket pose estimation. The MARNet consists of an encoder, a dual-pose-memory module (DPMM), and a decoder. The encoder extracts the global pose features from the input point clouds. The DPMM learns and memorizes the prototypical pose features covering all the samples by using the memory units at the training phase. Then, nondominant pose features are augmented by retrieving the stored prototypical pose features at the testing phase. At last, the decoder calculates the final poses. Our method can effectively alleviate the forgetting problem and strengthen the generalization by using the DPMM. The experimental results demonstrate that our method increases the estimation accuracy of rotation and translation by at least 1.49° and 0.021 m, respectively, for the bucket poses with small samples. This contributes to the development of intelligent mining excavators.

Novel multi-scale dilated CNN-LSTM for fault diagnosis of planetary gearbox with unbalanced samples under noisy environment

Lots of recent deep learning based intelligent fault diagnosis methods of planetary gearbox have achieved satisfactory accuracy with balanced training samples. Nevertheless, the fault samples are generally far less than healthy samples in practical engineering, and the collected data samples usually contain lots of noise, making it difficult to achieve accurate fault diagnosis. In order to solve these problems, this paper proposes a new method called novel multi-scale dilated convolutional neural network with long short-term memory (CNN-LSTM). Firstly, a novel multi-scale dilated CNN is constructed using new dilated strategy to enrich the coverage of the fields of view and avoid the loss of original information, which could adequately mine the distinguishing features of small samples. Secondly, an adaptive weight unit combined with LSTM is designed to fuse the distinguishing features and improve their robustness to noise. Finally, to pay more attention to the small samples and easily confused samples, a new-type loss function called enhanced cross entropy is developed. The test and analysis of the planetary gearbox data sets prove that the proposed method shows better diagnosis performance than other comparison methods using unbalanced training samples.

Transient stability evaluation model based on SSDAE with imbalanced correction

With the rapid development of machine learning technology, a new tool is provided for real-time stability evaluation in power systems. The training of a machine learning-based model is inseparable from a large number of training samples. However, compared with stable samples, unstable samples in power systems are infrequent. The results of the model evaluation are biased due to the imbalance of training samples. Faced with such a problem, a framework based on deep imbalanced learning is proposed. Firstly, for each sample, the samples nearby in the opposite class are applied to calculate its space information. Based on the space information of all samples, the spatial distribution characteristics of the training samples are obtained. And then, in order to obtain balanced training samples, unstable samples are generated according to their spatial distribution characteristics. Finally, stacked sparse denoising auto-encoder (SSDAE) based model, which has the ability of anti-noise, is established as the classifier. Simulation results in IEEE 39-bus system show the high performance of the proposed imbalanced correction scheme and evaluation scheme.

A data augmentation approach to train fully convolutional networks for left ventricle segmentation

Left ventricle (LV) segmentation plays an important role in the diagnosis of cardiovascular diseases. The cardiac contractile function can be quantified by measuring the segmentation results of LVs. Fully convolutional networks (FCNs) have been proven to be able to segment images. However, a large number of annotated images are required to train the network to avoid overfitting, which is a challenge for LV segmentation owing to the limited small number of available training samples. In this paper, we analyze the influence of augmenting training samples used in an FCN for LV segmentation, and propose a data augmentation approach based on shape models to train the FCN from a few samples. We show that the balanced training samples affect the performance of FCNs greatly. Experiments on four public datasets demonstrate that the FCN trained by our augmented data outperforms most existing automated segmentation methods with respect to several commonly used evaluation measures.

Integrating LiDAR and high-resolution imagery for object-based mapping of forest habitats in a heterogeneous temperate forest landscape

ABSTRACTLand-cover mapping (LCM) at a fine scale would be useful for forest management across heterogeneous natural landscapes. However, the heterogeneity of land covers at such scales results in complex spectral and textural properties that hinder the applicability of LCM. Besides, the method suffers from, e.g. inconsistent representation of different land-cover types, lack of sufficient and balanced training samples, and instability of classifiers trained by a high number of predictor variables. Even well-known object-based classification approaches are challenged with an objective evaluation of segmentation outputs. Here we classified partially ambiguous land-cover types across heterogeneous forest landscapes in the Bavarian Forest National Park (Germany) by combining metrics from airborne light detection and ranging (LiDAR) and colour infrared (CIR) imagery data and a random forest classifier implemented in an object-based paradigm. We evaluated the segmentation results by creating a global quality score based on inter- and intra-measurements of variance and the number of segments. Selected segmentation outputs were combined with balanced training samples to run the classification algorithm based on representative blocks within the national park. The entire processing chain was implemented in an open-source domain. The final segmentation consisted of LiDAR-based height, image-based Normalized Difference Vegetation Index (NDVI) and red band, with 20 cluster seeds and a minimum segment size of 40 pixels. In the classification, the most important variables included the height of the top layer, NDVI, Enhanced Vegetation Index (EVI) and Green–Red Vegetation Index (GRVI). The average values of 500 random forest runs indicated an overall accuracy of 86.6% and an estimated Cohen’s kappa coefficient of 85.2%, with different probabilities of correct classification for land-cover classes. Mature deciduous, standing deadwood, fallen deadwood, meadow, and bare soil classes were classified most accurately, whereas classification of young coniferous, intermediate-age coniferous, mature coniferous, young deciduous, and intermediate-age deciduous were associated with the highest uncertainties. Our methodology is sufficiently robust to be applied to other similarly structured sites across temperate forested landscapes. The versatility of the method is partially guaranteed by the proposed segmentation quality score, which satisfactorily corrects under- and over-segmentation.

Band Weighting via Maximizing Interclass Distance for Hyperspectral Image Classification

We present a novel band weighting strategy that exploits multiple binary support vector machines (SVMs) to maximize interclass spectral distances for multiclass hyperspectral remote image classification. Specifically, we commence by training binary SVMs based on the original training samples. We then balance the bands of training samples by maximizing the modified classification scores for SVMs. This balance scheme enlarges the distances between individual training samples and the SVM hyperplane. For each class, we reformulate the binary SVM objective function based on the balanced training samples, resulting in a weighting vector that associates a weight to each spectral band for the class. For a testing sample, we weight it and then classify it by using the binary SVM, both with respect to every individual class. The classification result is obtained from the classifier with the greatest score. Experiments on two benchmark data sets show the effectiveness of the proposed strategy.

Computer-aided lung nodule recognition by SVM classifier based on combination of random undersampling and SMOTE.

In lung cancer computer-aided detection/diagnosis (CAD) systems, classification of regions of interest (ROI) is often used to detect/diagnose lung nodule accurately. However, problems of unbalanced datasets often have detrimental effects on the performance of classification. In this paper, both minority and majority classes are resampled to increase the generalization ability. We propose a novel SVM classifier combined with random undersampling (RU) and SMOTE for lung nodule recognition. The combinations of the two resampling methods not only achieve a balanced training samples but also remove noise and duplicate information in the training sample and retain useful information to improve the effective data utilization, hence improving performance of SVM algorithm for pulmonary nodules classification under the unbalanced data. Eight features including 2D and 3D features are extracted for training and classification. Experimental results show that for different sizes of training datasets our RU-SMOTE-SVM classifier gets the highest classification accuracy among the four kinds of classifiers, and the average classification accuracy is more than 92.94%.

PAKDD'12 best paper

We consider the problem of generating balanced training samples from an unlabeled data set with an unknown class distribution. While random sampling works well when the data are balanced, it is ver...

PAKDD’12 best paper: generating balanced classifier-independent training samples from unlabeled data

We consider the problem of generating balanced training samples from an unlabeled data set with an unknown class distribution. While random sampling works well when the data are balanced, it is very ineffective for unbalanced data. Other approaches, such as active learning and cost-sensitive learning, are also suboptimal as they are classifier-dependent and require misclassification costs and labeled samples, respectively. We propose a new strategy for generating training samples, which is independent of the underlying class distribution of the data and the classifier that will be trained using the labeled data. Our methods are iterative and can be seen as variants of active learning, where we use semi-supervised clustering at each iteration to perform biased sampling from the clusters. We provide several strategies to estimate the underlying class distributions in the clusters and to increase the balancedness in the training samples. Experiments with both highly skewed and balanced data from the UCI repository and a private data set show that our algorithm produces much more balanced samples than random sampling or uncertainty sampling. Further, our sampling strategy is substantially more efficient than active learning methods. The experiments also validate that, with more balanced training data, classifiers trained with our samples outperform classifiers trained with random sampling or active learning.

Feature Ranking by Classification Accuracy Estimation of Multiple Data Samples

This article considers the gene ranking algorithm for the microarray data. The rank vector is estimated by classifications of the random data samples. At each iteration, the ranks of genes participating in the successful classification become higher. Unlike other methods of feature selection, the proposed algorithm allows increasing the generality of the classification models by construction of the balanced training samples and taking into account the descriptiveness of the gene combinations by the subset estimation.