Discriminative Regularization Research Articles

Discriminative Hamiltonian variational autoencoder for accurate tumor segmentation in data-scarce regimes

Deep learning has gained significant attention in medical image segmentation. However, the limited availability of annotated training data presents a challenge to achieving accurate results. In efforts to overcome this challenge, data augmentation techniques have been proposed. However, the majority of these approaches primarily focus on image generation. For segmentation tasks, providing both images and their corresponding target masks is crucial, and the generation of diverse and realistic samples remains a complex task, especially when working with limited training datasets. To this end, we propose a new end-to-end hybrid architecture based on Hamiltonian Variational Autoencoders (HVAE) and a discriminative regularization to improve the quality of generated images. Our method provides an accurate estimation of the joint distribution of the images and masks, resulting in the generation of realistic medical images with reduced artifacts and off-distribution instances. As generating 3D volumes requires substantial time and memory, our architecture operates on a slice-by-slice basis to segment 3D volumes, capitalizing on the richly augmented dataset. Experiments conducted on two public datasets, BRATS (MRI modality) and HECKTOR (PET modality), demonstrate the efficacy of our proposed method on different medical imaging modalities with limited data.

Adaptive Discriminative Regularization for Visual Classification

Adaptive Discriminative Regularization for Visual Classification

Multi-class feature selection via Sparse Softmax with a discriminative regularization

Multi-class feature selection via Sparse Softmax with a discriminative regularization

Locality Robust Domain Adaptation for cross-scene hyperspectral image classification

Domain adaptation (DA) has become a widely used technique for cross-scene hyperspectral image (HSI) classification. Most DA methods aim to learn a domain invariant subspace that reduce the domain discrepancy between source and target domains. However, some of them fail to explore the local manifold structure between different domains, while also neglecting the negative influences of abnormal features. To solve these problems, we propose a Locality Robust Domain Adaptation (LRDA) method for cross-domain data recognition. In LRDA, the statistical alignment is applied to reduce the domain-shift between the source and target domains. Then, LRDA combines the row-sparsity constraint and the discriminant regularization term to learn a robust projection matrix, while maintaining the discriminative capability of the matrix. Furthermore, a manifold regularization term is proposed to automatically learn the nearest neighbors and weights between two domains. The proposed LRDA not only reduces the discrepancy between two domains but also explores the local neighbor information between them. Experiment results on three HSI datasets illustrate that the proposed LRDA has better performance than other related methods.

Safe semi-supervised learning for pattern classification

Semi-supervised learning (SSL) based on manifold regularization in many fields has attracted widespread attention and research. However, SSL still has two main challenges: On the one hand, studies have shown that unlabeled data may cause performance degradation in semi-supervised classifiers, which means that unlabeled data introduces uncertainty and potential hazards. On the other hand, for samples distributed on different class boundaries, manifold regularization is not necessarily satisfactory, which will result in samples near the boundary that are likely to be misclassified. In response to the above problems, we propose a new SSL framework called safe semi-supervised learning (Sa-SSLJR for short). In Sa-SSLJR, a risk degree regularization term is constructed to estimate the uncertainty and potential risk of unlabeled data in the semi-supervised learning process. Secondly, based on manifold regularization and discriminant regularization, a joint regularization term is developed to solve the second challenge of SSL. Extensive experiments on multiple datasets show that our approach is competitive with state-of-the-art methods in terms of classification performance and feasibility.

Discriminative Projection Learning With Adaptive Reversed Graph Embedding for Supervised and Semi-Supervised Dimensionality Reduction

The graph embedding based dimensionality reduction (DR) methods have been widely used to overcome the curse of dimensionality problem in high-dimensional data. Especially in pattern recognition and data analysis, they are frequently used to extract features. However, there are two problems in the graph embedding based supervised DR methods: 1) the original data will inevitably be corrupted by noise or outliers, which makes the fixed graph constructed from these data unreliable; and 2) determining how to model the projection for enhancing the discrimination of projected samples. To overcome the above two challenges, this paper proposes a new supervised DR method called discriminative projection learning with adaptive reversed graph embedding (DP-ARGE). Specifically, DP-ARGE learns the graph dynamically in a low-dimensional subspace by using the low-dimensional embedding rather than projected samples for improving its robustness. In addition, a novel discriminative regularization term is designed to enhance the discrimination of the projection samples by using a few pairwise marginal heterogeneous samples rather than the all. To tackle the problem of estimating the neighborhood size parameter of the graph, a simple yet effective automatic parameter estimation strategy based on the density and similarity of data set is proposed. Furthermore, for better handling the data set with a few labels, DP-ARGE is extend to a semi-supervised method (SDP-ARGE) by a neighborhood preserving regularization term. SDP-ARGE can be used to preserve the local manifold structure of the data in the semi-supervised scenario. Comprehensive experiments show the superiority of the proposed methods in several real-world data sets.

A novel metric learning framework by exploiting global and local information

Distance metric learning plays a significant role in improving the generalization of algorithms related to distance metrics. In this paper, we first propose a generalized Mahalanobis metric learning framework (called GLML) to make use of the prior knowledge from the samples including global and local information, the main goal of which is to enlarge the distance of dissimilar sample pairs and shrink the distance of similar sample pairs simultaneously. Specifically, the metric learning problem is guided by a discriminative regularization through incorporating the pair-wise and class-wise information. Moreover, an effective alternating iterative algorithm is developed to optimize the proposed GLML, and the convergence of the algorithm is demonstrated theoretically. Then a boosting algorithm of GLML (BGLML) is proposed, where the low-rank basis learning is jointly optimized with the metric to better uncover the data structure and mitigate the computational burden. Following that, numerical experiments are carried out on binary classification and multi-classification problems. With different assessment criteria, experiment results from different scale datasets confirm that the proposed methods either improve generalization or have comparable results compared with the state-of-the-art algorithms.

Toward aircraft detection and fine-grained recognition from remote sensing images

The detection and recognition of aircraft on images is an important task in the field of optical remote sensing image interpretation. The challenges include the detection of arbitrary orientation targets and the fine-grained recognition of aircraft type. However, most related studies only pay attention to one of these because of the limitation of datasets, and an end-to-end network often cannot simultaneously solve these challenges well. Therefore, we propose a cascade framework for arbitrary-oriented and multitype aircraft detection in remote sensing images based on the convolutional neural network. A rotation-based detector that uses oriented boxes is designed to determine the position and direction of aircraft on remote sensing images. A fine-grained recognition subnetwork that adopts the ensemble learning and the Fisher discriminant regularization is designed to recognize the type of aircraft on images. The subnetwork is cascaded behind the rotation-based detector to achieve more accurate recognition. The mean average precision of this method reaches 83.7% on a challenging dataset, which is 2.4% to 8.2% higher than other advanced methods. Experiments demonstrated that the proposed method can greatly improve the accuracy of aircraft detection and recognition and that the method has good generalization performance.

Graph Convolutional Neural Networks with Geometric and Discrimination information

In recent years, geometric deep learning methods have been proposed, which are called Graph Convolutional Neural Networks (GCNNs). GCNNs not only can extract effective features like the classical CNN, but also can effectively reflect the true geometric structure of original data. Although GCNNs consider the geometric structure of original data, they construct the same feature graph to perform graph convolution, and ignore the difference between the local structures of different samples. Therefore, a novel Graph Convolutional Neural Network with Geometric and Discrimination information (GDGCNN) is proposed, which integrates traditional machine learning ideas to further improve the performance of feature extraction. In order to exploit differences between the local structures of different samples and make full use of the geometric structure of original data, GDGCNN constructs different feature graphs for different training batches to fully exploit the local geometry of data. Moreover, the discriminant regularization is introduced into GDGCNN to effectively utilize the discriminant information contained in original data. Therefore, GDGCNN has good discriminative ability and robustness. The experimental results show that GDGCNN can perform feature extraction tasks very well, and it is superior to some existing methods for classification in terms of accuracy and F1-Score.

Weighted Discriminative Sparse Representation for Image Classification

Sparse representation methods based on $$l _2$$ norm regularization have attracted much attention due to its low computational cost and competitive performance. How to enhance the discriminability of $$l _2$$ norm regularization-based representation method is a meaningful work. In this paper, we put forward a novel $$l _2$$ norm regularization-based representation method, called Weighted Discriminative Sparse Representation for Classification (WDSRC), in which we consider the global discriminability and the local discriminability using two discriminative regularization terms of representation. The global discriminability is obtained by decorrelating the representation results stemming from all distinct classes. The local discriminability is achieved by the weighted representation in which the representation coefficient of the training images dissimilar to the test image will be reduced and the representation coefficient of the training images similar to the test image will be increased, which restrains the training images dissimilar to the test image and promotes the training images similar to the test image as much as possible in representing the test sample. By considering the global and local discriminability of representations simultaneously, the proposed WDSRC method can gain more discriminative representation for classification. Extensive experiments on benchmark datasets of object, face, action and flower demonstrate the effectiveness of the proposed WDSRC method.

Geographical origin and level identification of frankincense based on hyperspectral image

Background: As the demand for traditional Chinese medicinal materials increases in China and even the world, there is an urgent need for an effective and simple identification technology to identify the origin and quality of the latter and ensure the safety of clinical medication. Mineral element analysis and isotope finger-printing are the two commonly used techniques in traditional origin identification. Both of these techniques require the use of stoichiometric methods in the identification process. Although they have high accuracy and sensitivity, they are expensive and inefficient. In addition, near-infrared spectroscopy is a fast, nondestructive, and widely used identification technique developed in recent years, but its identification results are susceptible to samples' states and environmental conditions, and its sensitivity is low. Hyperspectral imaging combines the advantages of imaging technology and optical technology, which can simultaneously access the image information and spectral information which reflect the external characteristics, internal physical structure, and chemical composition of the samples. Hyperspectral imaging is widely applied to agricultural product inspection, but research into its application in origin and quality identification of TCM materials is rare. Methods: In this study, the algorithm framework discriminative marginalized least squares regression (DMLSR) was used for feature extraction of frankincense hyperspectral data. The DMLSR with intraclass compactness graph and manifold regularization can efficiently learn the projective samples with higher separability and less redundant information than the original samples. Then, the discriminative collaborative representation with Tikhonov regularization (DCRT) was applied for classifying the geographical origin and level of frankincense. DCRT introduces the discriminant regularization term and incorporates SID, which is more sensitive to the spectrum as the measurement method and is more suitable for the frankincense spectral data compared with SVM. Results: For the origin classification task, samples of all levels from each origin were, respectively, selected for three-way classification. We used 10-fold cross-validation to select a model parameter in the experiment. When obtaining the optimal parameters, we randomly selected the training set and testing set, where the training set accounts for 70% and the training set for 30%. After repeating this random process 10 times, we obtained the final average classification accuracy, which is higher than 90%, and the standard deviation fluctuation is usually small. For the level classification task, samples of each level from three origins were separately selected for multiclassification. We randomly selected the training set and testing set from each origin. The level classification results of the three origins are good on D4350 data, and the classification accuracy of each level is basically above 80%. Conclusion: Experiments and analysis show that our algorithm framework has excellent classification performance, which is stable in origin classification and has potential for generalization. In addition, the experiments show that in our algorithm framework, different classification tasks need to combine different data sources to achieve better classification and recognition, as the origin classification task uses frankincense's D3000 data, and level classification task uses frankincense's D4350 data.

Class-specific discriminant regularization in real-time deep CNN models for binary classification problems

In this paper, we first propose lightweight deep CNN models, capable of effectively operating in real-time on-drone for high-resolution video input, addressing various binary classification problems, e.g. crowd, face, football player, and bicycle detection, in the context of media coverage of specific sport events by drones with increased decisional autonomy. Furthermore, we propose a novel class-specific discriminant regularizer in order to improve the generalization ability of the proposed real-time models, exploiting the nature of the considered two-class problems. The experimental evaluation on four datasets validates the effectiveness of the proposed regularizer in enhancing the generalization ability of the proposed models.

Robust Discriminative Metric Learning for Image Representation

Metric learning has attracted significant attention in the past decades, because of its appealing advances in various real-world tasks, e.g., person re-identification and face recognition. Traditional supervised metric learning attempts to seek a discriminative metric, which could minimize the pairwise distance of within-class data samples, while maximizing the pairwise distance of data samples from various classes. However, it is still a challenge to build a robust and discriminative metric, especially for corrupted data in the real-world application. In this paper, we propose a Robust Discriminative Metric Learning algorithm through fast low-rank representation and denoising strategy. To be specific, the metric learning problem is guided by a discriminative regularization by incorporating the pair-wise or class-wise information. Moreover, the low-rank basis learning is jointly optimized with the metric to better uncover the global data structure and remove noise. Furthermore, the fast low-rank representation is implemented to mitigate the computational burden and ensure the scalability on large-scale datasets. Finally, we evaluate our learned metric on several challenging tasks, e.g., face recognition/verification, object recognition, image clustering, and person re-identification. The experimental results verify the effectiveness of our proposed algorithm in comparison to many metric learning algorithms, even deep learning ones.

Supervised cross-fusion method: a new triplet approach to fuse thermal, radar, and optical satellite data for land use classification

This study presents a new fusion method namely supervised cross-fusion method to improve the capability of fused thermal, radar, and optical images for classification. The proposed cross-fusion method is a combination of pixel-based and supervised feature-based fusion of thermal, radar, and optical data. The pixel-based fusion was applied to fuse optical data of Sentinel-2 and Landsat 8. According to correlation coefficient (CR) and signal to noise ratio (SNR), among the used pixel-based fusion methods, wavelet obtained the best results for fusion. Considering spectral and spatial information preservation, CR of the wavelet method is 0.97 and 0.96, respectively. The supervised feature-based fusion method is a fusion of best output of pixel-based fusion level, land surface temperature (LST) data, and Sentinel-1 radar image using a supervised approach. The supervised approach is a supervised feature selection and learning of the inputs based on linear discriminant analysis and sparse regularization (LDASR) algorithm. In the present study, the non-negative matrix factorization (NMF) was utilized for feature extraction. A comparison of the obtained results with state of the art fusion method indicated a higher accuracy of our proposed method of classification. The rotation forest (RoF) classification results improvement was 25% and the support vector machine (SVM) results improvement was 31%. The results showed that the proposed method is well classified and separated four main classes of settlements, barren land, river, river bank, and even the bridges over the river. Also, a number of unclassified pixels by SVM are very low compared to other classification methods and can be neglected. The study results showed that LST calculated using thermal data has had positive effects on improving the classification results. By comparing the results of supervised cross-fusion without using LST data to the proposed method results, SVM and RoF classifiers showed 38% and 7% of classification improvement, respectively.

Learning Compact and Discriminative Stacked Autoencoder for Hyperspectral Image Classification

As one of the fundamental research topics in remote sensing image analysis, hyperspectral image (HSI) classification has been extensively studied so far. However, how to discriminatively learn a low-dimensional feature space, in which the mapped features have small within-class scatter and big between-class separation, is still a challenging problem. To address this issue, this paper proposes an effective framework, named compact and discriminative stacked autoencoder (CDSAE), for HSI classification. The proposed CDSAE framework comprises two stages with different optimization objectives, which can learn discriminative low-dimensional feature mappings and train an effective classifier progressively. First, we impose a local Fisher discriminant regularization on each hidden layer of stacked autoencoder (SAE) to train discriminative SAE (DSAE) by minimizing reconstruction error. This stage can learn feature mappings, in which the pixels from the same land-cover class are mapped as nearly as possible and the pixels from different land-cover categories are separated by a large margin. Second, we learn an effective classifier and meanwhile update DSAE with a local Fisher discriminant regularization being embedded on the top of feature representations. Moreover, to learn a compact DSAE with as small number of hidden neurons as possible, we impose a diversity regularization on the hidden neurons of DSAE to balance the feature dimensionality and the feature representation capability. The experimental results on three widely-used HSI data sets and comprehensive comparisons with existing methods demonstrate that our proposed method is effective.

Fisher-discriminative regularized latent sparse transfer model

Top performing algorithms are trained on massive amounts of labeled data. Alternatively, domain adaptation (DA) provides an attractive way to address the few labeled tasks when the labeled data from a different but related domain are available. Motivated by Fisher criterion, we present the novel discriminative regularization term on the latent subspace which incorporates the latent sparse domain transfer (LSDT) model in a unified framework. The key underlying idea is to make samples from one class closer and farther away from different class samples. However, it is nontrivial to design the efficient optimization algorithm. Instead, we construct a convex surrogate relaxation optimization constraint to ease this issue by alternating direction method of multipliers (ADMM) algorithm. Subsequently, we generalize our model in the reproduced kernel Hilbert space (RKHS) for tracking the nonlinear domain shift. Empirical studies demonstrate the performance improvement on the benchmark vision dataset Caltech-4DA.

Weighted feature selection via discriminative sparse multi-view learning

The matrix-based structured sparsity-inducing multi-view feature selection has received much attention because it can select the relevant features through the information-rich multi-view data instead of the single-view data. In this paper, a novel supervised sparse multi-view feature selection model is proposed based on the separable weighted loss term and the discriminative regularization terms. The proposed model adopts the separable strategy to enforce the weighted penalty for each view instead of using the concatenated feature vectors to calculate the penalty. Therefore, the proposed model is established by considering both the complementarity of multiple views and the specificity of each view. The derived model can be split into several small-scale problems in the process of optimization, and be solved efficiently via an iterative algorithm with low complexity. Furthermore, the convergence of the proposed iterative algorithm is investigated from both theoretical and experimental aspects. The extensive experiments compared with several state-of-the-art matrix-based feature selection methods on the widely used multi-view datasets show the effectiveness of the proposed method.

Discriminative regularization of the latent manifold of variational auto-encoders

We present an approach on training classifiers or regressors using the latent embedding of variational auto-encoders (VAE), an unsupervised deep learning method, as features. Usually VAEs are trained using unlabeled data and independently from the classifier, whereas we investigate and analyze the performance of a classifier or regressor that is trained jointly with the variational deep network. We found that models trained this way can improve the embedding s.t. to increase classification performance, and also can be used for semi-supervised learning, building up the information extracting latent representation in an incremental fashion.The model was tested on two widely known computer vision benchmarks, and its generalization power was evaluated on an independent dataset. Additionally, generally applicable statistical methods are presented for evaluating similarly performing classifiers, and used to quantify the performance increase. The general applicability and ease-of-use of deep learning approaches allows for a wide applicability of the method.

Supervised feature extraction method based on low-rank representation with preserving local pairwise constraints for hyperspectral images

Feature extraction (FE) methods based on low-rank representation (LRR) have become important topics in hyperspectral images (HSIs) data analysis. In this paper, a supervised FE method for HSIs data based on LRR with the ability to preserve the local pairwise constraints information (LRLPC) is proposed. LRLPC does not change the data dimensionality and only employs a technique to enrich the original feature space (OFS) and to obtain enriched feature space, which results in features richer than OFS. To overcome the problem of LRR in lacking the local structure information (LSI) of data, a local discriminative regularization term is imposed on the fitness function of LRR to keep the LSI of data. For nonlinear structure of data, LRLPC is extended to kernel LRLPC (KLRLPC) using kernel trick. Utilization of existing information in the pairwise constraints is useful for limited labeled samples situations as a common problem in HSI data analysis. The obtained experimental results using two well-known HSI data sets confirm the effectiveness of LRLPC and KLRLPC for dimension reduction and classification of HSIs.

Hyperspectral Image Classification With Discriminative Kernel Collaborative Representation and Tikhonov Regularization

Recently, collaborative representation has received much attention in the hyperspectral image (HSI) classification due to its simplicity and effectiveness. However, the existing collaborative representation-based HSI classification methods ignore the correlation among different classes. To overcome this problem, we propose a discriminative kernel collaborative representation and Tikhonov regularization method (DKCRT) for HSI classification, which can make the kernel collaborative representation of different classes to be more discriminative. Specifically, the kernel trick is adopted to map the original HSI into a high space to improve the class separability. Besides, distance-weighted kernel Tikhonov regularization is adopted to enforce these training samples to have large representation coefficients, which are similar to the test sample in the high-dimensional feature space. Moreover, we add a discriminative regularization term to further enhance the separability of different classes, which can take the correlation among different classes into consideration. Furthermore, to take the spatial information of HSI into consideration, we extend the DKCRT to a joint version named JDKCRT. Experiments on real HSIs demonstrate the efficiency of the proposed DKCRT and JDKCRT.