Open Set Recognition and Category Discovery Framework for SAR Target Classification Based on K-Contrast Loss and Deep Clustering

The central obstacle in Open Set Recognition (OSR) is striking a balance between minimizing classification errors on known data and managing the risks posed by open space for unknown data. To address these issues, we present three novel frameworks: the Positive–Negative Prototypes Fusion Framework (PNPFF), its adversarial extension (APNPFF), and an enhanced version, APNPFF++. The PNPFF framework incorporates multiple positive prototypes to capture intra-class variability and a single negative prototype to strengthen intra-class compactness and inter-class separation. This approach reduces the classification risk for known data while reserving space for unknowns, partially alleviating open space risk. Building on this, APNPFF and APNPFF++ use Generative Adversarial Network (GAN) and manifold mix-up methods, respectively, to simulate two parts of unknown class data, further reducing open space risk and enhancing the model’s generalization performance. Comprehensive experiments conducted across multiple benchmark datasets demonstrate the effectiveness and robustness of the proposed methods, highlighting their superiority in handling both known and unknown class recognition tasks.

In the unsupervised open set domain adaptation (UOSDA), the target domain contains unknown classes that are not observed in the source domain. Researchers in this area aim to train a classifier to accurately: 1) recognize unknown target data (data with unknown classes) and 2) classify other target data. To achieve this aim, a previous study has proven an upper bound of the target-domain risk, and the open set difference, as an important term in the upper bound, is used to measure the risk on unknown target data. By minimizing the upper bound, a shallow classifier can be trained to achieve the aim. However, if the classifier is very flexible [e.g., deep neural networks (DNNs)], the open set difference will converge to a negative value when minimizing the upper bound, which causes an issue where most target data are recognized as unknown data. To address this issue, we propose a new upper bound of target-domain risk for UOSDA, which includes four terms: source-domain risk, ϵ -open set difference ( ∆ϵ ), distributional discrepancy between domains, and a constant. Compared with the open set difference, ∆ϵ is more robust against the issue when it is being minimized, and thus we are able to use very flexible classifiers (i.e., DNNs). Then, we propose a new principle-guided deep UOSDA method that trains DNNs via minimizing the new upper bound. Specifically, source-domain risk and ∆ϵ are minimized by gradient descent, and the distributional discrepancy is minimized via a novel open set conditional adversarial training strategy. Finally, compared with the existing shallow and deep UOSDA methods, our method shows the state-of-the-art performance on several benchmark datasets, including digit recognition [modified National Institute of Standards and Technology database (MNIST), the Street View House Number (SVHN), U.S. Postal Service (USPS)], object recognition (Office-31, Office-Home), and face recognition [pose, illumination, and expression (PIE)].

In recent years, synthetic aperture radar automatic target recognition (SAR ATR) has been widely researched for its ability to achieve high-performance target classification through supervised training, facilitating battlefield reconnaissance, and intelligence generation. When dealing with unknown class data for ATR model training, significant time and effort are typically required for manual interpretation and labeling. However, when unknown class data shares the same domain as the known labeled data in the library, leveraging their shared deep semantic knowledge can enable automatic classification and labeling of the unknown class data. In this article, we investigate the novel category discovery (NCD) problem in SAR images, using labeled data to guide the clustering process of new class data. Specifically, we utilize the Unified Objective function training framework to address the training imbalance between labeled and unlabeled data in the NCD process, incorporating various improvements on this foundation. Through a binary segmentation-based strategy, we effectively mitigate the interference of “noise pairs” with significant semantic differences on the model's self-supervised pretraining. In addition, we introduce multicrop consistency loss and equal distance loss to impose constraints on training by leveraging intraclass and interclass relationships in the latent space, thereby obtaining representations with higher interclass separability. Our method achieves state-of-the-art clustering performance in multiple scenarios. Extensive experimental results on the MSTAR benchmark dataset demonstrate the effectiveness of the proposed methods.

This paper presents our automated filter-synthesis algorithm for the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). We discuss use of this autoMinace filter in face recognition and automatic target recognition (ATR), in which we consider both true-class object classification and rejection of non-database objects (impostors in face recognition and confusers in ATR). We use at least one Minace filter per object class to be recognized; a separate Minace filter or a set of Minace filters is synthesized for each object class. The Minace parameter c trades-off distortion-tolerance (recognition) versus discrimination (impostor/confuser/clutter rejection) performance. Our automated Minace filter-synthesis algorithm (autoMinace) automatically selects the Minace filter parameter c and selects the training set images to be included in the filter, so that we achieve both good recognition and good impostor/confuser and clutter rejection performance; this is achieved using a training and validation set. No impostor/confuser, clutter or test set data is present in the training or validation sets. Use of the peak-to-correlation energy (PCE) ratio is found to perform better than the correlation peak height metric. The use of circular versus linear correlations is addressed; circular correlations require less storage and fewer online computations and are thus preferable. Representative test results for three different databases - visual face, IR ATR, and SAR ATR - are presented. We also discuss an efficient implementation of Minace filters for detection applications, where the filter template is much smaller than the input target scene.

Spatial Location Constraint Prototype Loss for Open Set Recognition

Spatial Location Constraint Prototype Loss for Open Set Recognition

Conditional feature generation for transductive open-set recognition via dual-space consistent sampling

Image classification methods are usually trained to perform predictions taking into account a predefined group of known classes. Real-world problems, however, may not allow for a full knowledge of the input and label spaces, making failures in recognition a hazard to deep visual learning. Open set recognition methods are characterized by the ability to correctly identify inputs of known and unknown classes. In this context, we propose GeMOS: simple and plug-and-play open set recognition modules that can be attached to pretrained Deep Neural Networks for visual recognition. The GeMOS framework pairs pre-trained Convolutional Neural Networks with generative models for open set recognition to extract open set scores for each sample, allowing for failure recognition in object recognition tasks. We conduct a thorough evaluation of the proposed method in comparison with state-of-the-art open set algorithms, finding that GeMOS either outperforms or is statistically indistinguishable from more complex and costly models.

In Synthetic Aperture Radar (SAR) image classification tasks, the performance depends on both the classifier and the dataset itself. However, in comparison with plenty of SAR classification methods, there is little work aimed at analyzing the distinguishability of the dataset. In the classification dataset, some classes are semantically different but their distinguishability is low, the classes are hard to be classified especially in some more practical cases that there are unknown classes without supervision exist. Referring to open set recognition (OSR), in this paper, we proposed the SAR Distinguishability Analysor (SAR-DA) to evaluate the distinguishability of the OpenSARUrban dataset. By modeling each class as a multivariate Gaussian distribution in latent space, SAR-DA can not only classify the classes having been seen in training phase, but also can recognize unknown samples if a test sample is out of each known distribution. Each class in OpenSARUr-ban is set unknown in turn, then we apply the SAR-DA on the split dataset in OSR and supervised setting. The distinguishability can be reflected by the unknown recognition recall rate. The experimental results show that the unknown recognition recall rate in OSR setting significantly decreased compared with those in supervised setting, indicating that even though the classes in OpenSARUrban are semantically different from each other, the latent distributions of some classes are quite similar and hard to be classified, thus these classes are of low distinguishability.

Few-shot open-set recognition (FSOSR) poses a significant challenge as it requires identifying unknown classes while maintaining the classification performance of known classes, despite having limited access to labeled training samples. Current methods often employ non-directional metric-based losses to encapsulate feature attributes within the embedding space, inadvertently disregarding the potential influence of spatial distribution deviations of feature representations on open-set recognition performance. To address this, we present a novel directional metric-based method termed Bilevel Direction Preserving (BiDirP). This method incorporates two direction-preserving regularizers operating at distinct levels, specifically at the instance and prototype levels. The combined application of these two direction-preserving regularizers effectively enhances the spatial separation between prototypes of different classes and refines the classification decision boundaries, which results in an improved discriminative ability to differentiate unknown classes within a broader open space. Comprehensive experiments on public benchmarks show that BiDirP can significantly improve the detection ability of unknown classes while correctly classifying known classes.

Understanding open-set recognition by Jacobian norm and inter-class separation

Open set recognition requires models to recognize samples of known classes learned in the training set while reject unknowns not learned. Compared with the structural risk minimization theory for closed-set problems, structural risk in open set tasks remains rarely explored. In this paper, we point out that balancing between structural risk and open space risk is crucial for open set recognition, and re-formalize it as open set structural risk. This brings a new view towards the general relationship between closed set recognition and open set recognition against the common intuition, which argues that a good closed set classifier always benefits for open set recognition. Specifically, we theoretically and experimentally show that recent mix-based data augmentation methods are aggressive closed set regularization methods, which reduce structural risk at cost of sacrificing open space risk. Besides, we show that existing negative data augmentation designed for open space risk reduction also ignore the trade-off problem between structural risk and open space risk, which limits their performance. We propose an efficient negative data augmentation strategy named self-mix and a corresponding method named OpenMix. OpenMix generates high-quality negative samples by mixing samples themselves, which can take care of both risks simultaneously. When combining OpenMix with conservative closed set regularization methods to form OpenMix+, models can achieve lower open set structural risk. Extensive experiments validate the superiority of OpenMix and OpenMix+ in terms of both effectiveness and universality.

In order to improve the performance of hyperspectral open set recognition processing with insufficient sample size, the authors propose a new supervised contrastive learning (SCL) framework (FSSCL‐OSC) that can achieve open‐set classification of hyperspectral images in scenarios with very few sample sizes, which consists of three steps. First, supervised contrastive learning (SCL) with geometric transformations is designed, which uses rotated labels as supervision to obtain lower‐level characteristics that more accurately capture various orientations and then makes use of a spectral‐spatial characteristic extraction network to maximize the utilization of HSI's spectrum and spatial information. Next, a class anchor open‐set classification module based on a twin clustering network is designed, which uses SCL and FSL to extract more specific personal information by mining the category‐invariant characteristics present in the known versus unknown class data. Finally, multiple convolution and open‐set recognition (OSC) is performed on the feature blocks. Experimental results on three classical HSI datasets show that FSSCL‐OSC provides a significant improvement over existing methods, under a sample size of only 10%, the overall accuracy reached 82.38%, 90.76% and 84.70%, respectively.

Open set recognition (OSR) in radio frequency fingerprint (RFF) is critical for securing Internet of Things (IoT) systems, where previously unseen or malicious devices may attempt unauthorized access. A widely used approach treats all unknown devices as a single additional class and assumes that they will produce low confidence scores during inference. However, due to the inherently subtle and highly similar RFF features across devices, this assumption often fails, leading to high false acceptance rates. To address this challenge, we propose a novel framework, called Multiple Unknown Classes Generation (MUCG), which replaces the single-class modeling of unknowns with a more expressive structure that simulates multiple distinct unknown classes. MUCG employs a conditional generative adversarial network (cGAN) guided by ideal signal priors to produce diverse and realistic unknown samples. Furthermore, we introduce a soft label perturbation (SLP) strategy that blends label semantics using Feature-wise Linear Modulation (FiLM), encouraging the generator to embed richer feature variations. Experiments on three public IoT datasets demonstrate that MUCG consistently outperforms state-of-the-art (SOTA) methods in OSR tasks, achieving superior accuracy and robustness under varying signal conditions.

This paper proposes a method to use deep neural networks as end-to-end open-set classifiers. It is based on intra-class data splitting. In open-set recognition, only samples from a limited number of known classes are available for training. During inference, an open-set classifier must reject samples from unknown classes while correctly classifying samples from known classes. The proposed method splits given data into typical and atypical normal subsets by using a closed-set classifier. This enables to model the abnormal classes by atypical normal samples. Accordingly, the open-set recognition problem is reformulated into a traditional classification problem. In addition, a closed-set regularization is proposed to guarantee a high closed-set classification performance. Intensive experiments on five well-known image datasets showed the effectiveness of the proposed method which outperformed the baselines and achieved a distinct improvement over the state-of-the-art methods.