General Feature Space Research Articles

Vocoder Detection of Spoofing Speech Based on GAN Fingerprints and Domain Generalization

As an important part of the text-to-speech (TTS) system, vocoders convert acoustic features into speech waveforms. The difference in vocoders is key to producing different types of forged speech in the TTS system. With the rapid development of general adversarial networks (GANs), an increasing number of GAN vocoders have been proposed. Detectors often encounter vocoders of unknown types, which leads to a decline in the generalization performance of models. However, existing studies lack research on detection generalization based on GAN vocoders. To solve this problem, this study proposes vocoder detection of spoofed speech based on GAN fingerprints and domain generalization. The framework can widen the distance between real speech and forged speech in feature space, improving the detection model’s performance. Specifically, we utilize a fingerprint extractor based on an autoencoder to extract GAN fingerprints from vocoders. We then weight them to the forged speech for subsequent classification to learn the forged speech features with high differentiation. Subsequently, domain generalization is used to further improve the generalization ability of the model for unseen forgery types. We achieve domain generalization using domain-adversarial learning and asymmetric triplet loss to learn a better generalized feature space in which real speech is compact and forged speech synthesized by different vocoders is dispersed. Finally, to optimize the training process, curriculum learning is used to dynamically adjust the contributions of the samples with different difficulties in the training process. Experimental results show that the proposed method achieves the most advanced detection results among four GAN vocoders. The code is available at https://github.com/multimedia-infomation-security/GAN-Vocoder-detection .

Discrimination between icequakes and earthquakes in southern Alaska: an exploration of waveform features using Random Forest algorithm

SUMMARY This study examines the feature space of seismic waveforms often used in machine learning applications for seismic event detection and classification problems. Our investigation centres on the southern Alaska region, where the seismic record captures diverse seismic activity, notably from the calving of marine-terminating glaciers and tectonic earthquakes along active plate boundaries. While the automated discrimination of earthquakes and glacier quakes is our nominal goal, this data set provides an outstanding opportunity to explore the general feature space of regional seismic phases. That objective has applicability beyond ice quakes and our geographic region of study. We make a noteworthy discovery that features rooted in the spectral content of seismic waveforms consistently outperform statistical and temporal features. Spectral features demonstrate robust performance, exhibiting resilience to class imbalance while being minimally impacted by factors such as epicentral distance and signal-to-noise ratio. We also conduct experiments on the transferability of the model and find that transferability primarily depends on the appearance of the waveforms. Finally, we analyse misclassified events and find examples that are identified incorrectly in the original regional catalogue.

PKRT-Net: Prior knowledge-based relation transformer network for optic cup and disc segmentation

Glaucoma causes irreversible vision loss, and early detection of glaucoma is essential to protect the vision of patients. The optic cup (OC) and optic disc (OD) are two critical anatomical structures for glaucoma diagnosis. Methods based on convolutional neural networks (CNNs) have been proposed to extract OC and OD, in which OC extraction is very challenging. However, the clinical prior knowledge is not fully utilized in existing CNN methods, which limits the performance of extracting OC and OD. Besides, CNN methods cannot learn long-range semantic information interaction well due to the intrinsic locality of convolution operations. In this paper, we propose a Prior Knowledge-based Relation Transformer Network (PKRT-Net), which employs the clinical prior knowledge to assist OC segmentation and model efficient long-range relation of spatial features by the transformer. PKRT-Net consists of a dual-branch module, a relation transformer fusion module, and a decoder with weighted fusion. Dual-branch module decouples the fundus image into the vessel feature space and general local feature space; the relation transformer fusion module fuses the clinical prior information with local features to obtain more representative features; the weighted fusion module fuses the multi-scale side-outputs from the decoder with the representation of relation transformer module to improve the segmentation performance. We evaluate our proposed PKRT-Net on three public available OC and OD segmentation datasets (i.e., Drishti-GS, RIM-ONE(r3), and REFUGE). The experimental results demonstrate that our proposed PKRT-Net framework achieves state-of-the-art OC and OD segmentation results on these three public datasets.

Domain knowledge aided machine learning method for properties prediction of soft magnetic metallic glasses

A machine learning (ML) method aided by domain knowledge was proposed to predict saturated magnetization (Bs) and critical diameter (Dmax) of soft magnetic metallic glasses (MGs). Two datasets were established based on published experimental works about soft magnetic MGs. A general feature space was proposed and proven to be adaptive for ML model training for different prediction tasks. It was demonstrated that the predictive performance of ML models was better than that of traditional knowledge-based estimation methods. In addition, domain knowledge aided feature design can greatly reduce the number of features without significantly reducing the prediction accuracy. Finally, the binary classification of Dmax of soft magnetic MGs was studied.

Parametric model embedding

Methodologies for reducing the design-space dimensionality in shape optimization have been recently developed based on unsupervised machine learning methods. These methods provide reduced dimensionality representations of the design space, capable of maintaining a certain degree of the original design variability. Nevertheless, they usually do not allow to use directly the original parameterization method, representing a limitation to their widespread application in the industrial field, where the design parameters often pertain to well-established parametric models, e.g. CAD (computer aided design) models. This work presents how to embed the parametric-model original parameters in a reduced-dimensionality representation of the design space. The method, which takes advantage from the definition of a newly-introduced generalized feature space, is demonstrated, as a proof of concept, for the reparameterization of 2D Bezier curves and 3D free-form deformation design spaces and the consequent solution of simulation-driven design optimization problems of a subsonic airfoil and a naval destroyer in calm water, respectively.

Adversarial learning and decomposition-based domain generalization for face anti-spoofing

Face anti-spoofing (FAS) plays a critical role in the face recognition community for securing the face presentation attacks. Many works have been proposed to regard FAS as a domain generalization problem for robust deployment in real-world scenarios. However, existing methods focus on extracting intrinsic spoofing cues to improve the generalization ability, yet neglect to train a robust classifier. In this paper, we propose a framework to improve the generalization ability of face anti-spoofing in two folds:) a generalized feature space is obtained via aggregation of all live faces while dispersing each domain’s spoof faces; and) a domain agnostic classifier is trained through low-rank decomposition. Specifically, a Common Specific Decomposition for Specific (CSD-S) layer is deployed in the last layer of the network to select common features while discarding domain-specific ones among multiple source domains. The above-mentioned two components are integrated into an end-to-end framework, ensuring the generalization ability to unseen scenarios. The extensive experiments demonstrate that the proposed method achieves state-of-the-art results on four public datasets, including CASIA-MFSD, MSU-MFSD, Replay-Attack, and OULU-NPU.

Fast Ship Detection With Spatial-Frequency Analysis and ANOVA-Based Feature Fusion

High-frequency surface wave radar (HFSWR) can be effectively used to detect ships in the exclusive economic zone. However, the ship signal is concealed and interfered with various clutter and background noise in the Doppler spectrum. In this letter, a range-Doppler (RD) image-based novel ship detection algorithm is proposed by exploiting spatial-frequency information and a unique feature fusion based on the analysis of variance. The algorithm subsumes three successive stages: Stage I—the plausible region of interest is captured, Stage II—the features from different sources are fused into one generalized feature space, and Stage III—an extreme learning machine-based classifier is utilized to localize the ships. Experimental results on challenging HFSWR-RD datasets demonstrate that the proposed algorithm has a competitive performance over other ship detection algorithms.

Selective Domain-Invariant Feature Alignment Network for Face Anti-Spoofing

One primary challenge in face anti-spoofing refers to suffering a sharp performance drop in cross-domain scenes, where training and testing images are collected from different datasets. Recent methods have achieved promising results by aligning the features of all images among the available source domains. However, due to significant distribution discrepancies among non-face regions of all images, it is challenging to capture domain-invariant features for these regions. In this paper, we propose a novel Selective Domain-invariant Feature Alignment Network (SDFANet) for cross-domain face anti-spoofing, which aims to seek common feature representations by fully exploring the generalization of different regions of images. Different from previous works that align the whole features directly, the proposed SDFANet leverages multiple domain discriminators with the same architecture to balance the generalization of different regions of the all images. Specifically, we firstly design a multi-grained feature alignment network composed of a local-region and global-image alignment subnetworks to learn more generalized feature space for real faces. Besides, the domain adapter module, which aims to alleviate the large domain discrepancy with the help of the domain attention strategy, is adopted to facilitate the learning of our multi-grained feature alignment network. In addition, a multi-scale attention fusion module is designed in our feature generator to refine the different levels of features effectively. Experimental results show that the proposed SDFANet can greatly improve the generalization ability of face anti-spoofing, and that is superior to the existing methods.

Fast and robust visual tracking with hard balanced focal loss and guided domain adaption

Recently, Siamese networks based trackers have shown excellent performance in accuracy and speed. However, previous studies treat all training samples equally and use the general feature space without adapting to the specific video during tracking. These trackers ignore the class data imbalance during training and the feature space difference between the generic domain and the current tracking target domain, which limits the robustness of trackers. In this paper, we propose an algorithm for learning a discriminative and self-adaptive feature representation, in order to achieve accurate and robust tracking. During the off-line training stage, a hard balanced focal loss function is utilized to solve the positive–negative samples imbalance and the hard-easy negative samples imbalance. During the tracking phase, an off-line trained guided domain adaptation module is embedded into the Siamese networks, which can quickly transfer the feature space from the general domain to the current video domain by adjusting the search branch channel weights. Our networks are trained in an end-to-end manner and without online updating. Our tracker runs at 130 FPS while achieving favorable performance against the state-of-the-art methods on OTB-2013, OTB-2015, VOT-2016, VOT-2017, GOT-10 K and TC-128 benchmarks.

A Width-growth Model with Subnetwork Nodes and Refinement Structure for Representation Learning and Image Classification

This article presents a new supervised multilayer subnetwork-based feature refinement and classification model for representation learning. The novelties of this algorithm are as follows: 1) different from most multilayer networks that go deeper with increased number of network layers, this work architects a model with wider subnetwork nodes; 2) the conventional classification methods adopt a separate search mechanism to derive a generalized feature space and to get the final cognition, but this work proposes a one-shot process to find the meaningful latent space and recognize the objects; and 3) the traditional feature representation and image classification approaches apply a unimodal feature coding, which suffers from lack of global knowledge. This work overcomes the pitfall through multimodal fusion that fuses various feature sources into one superstate encoding to achieve higher performance. A cross-domain experimental study on camera identification and image classification shows that the proposed method achieves superior performance compared to the existing models.

Learning of Planning Models for Dexterous Manipulation Based on Human Demonstrations

In the human environment service robots have to be able to manipulate autonomously a large variety of objects in a workspace restricted by collisions with obstacles, self-collisions and task constraints. Planning enables the robot system to generalize predefined or learned manipulation knowledge to new environments. For dexterous manipulation tasks the manual definition of planning models is time-consuming and error-prone. In this work, planning models for dexterous tasks are learned based on multiple human demonstrations using a general feature space including automatically generated contact constraints, which are automatically relaxed to consider the correspondence problem. In order to execute the learned planning model with different objects, the contact location is transformed to given object geometry using morphing. The initial, overspecialized planning model is generalized using a previously described, parallelized optimization algorithm with the goal to find a maximal subset of task constraints, which admits a solution to a set of test problems. Experiments on two different, dexterous tasks show the applicability of the learning approach to dexterous manipulation tasks.

A general feature space for automatic verb classification

Lexical semantic classes of verbs play an important role in structuring complex predicate information in a lexicon, thereby avoiding redundancy and enabling generalizations across semantically similar verbs with respect to their usage. Such classes, however, require many person-years of expert effort to create manually, and methods are needed for automatically assigning verbs to appropriate classes. In this work, we develop and evaluate a feature space to support the automatic assignment of verbs into a well-known lexical semantic classification that is frequently used in natural language processing. The feature space is general – applicable to any class distinctions within the target classification; broad – tapping into a variety of semantic features of the classes; and inexpensive – requiring no more than a POS tagger and chunker. We perform experiments using support vector machines (SVMs) with the proposed feature space, demonstrating a reduction in error rate ranging from 48% to 88% over a chance baseline accuracy, across classification tasks of varying difficulty. In particular, we attain performance comparable to or better than that of feature sets manually selected for the particular tasks. Our results show that the approach is generally applicable, and reduces the need for resource-intensive linguistic analysis for each new classification task. We also perform a wide range of experiments to determine the most informative features in the feature space, finding that simple, easily extractable features suffice for good verb classification performance.

The Two-Point Correlation Function: A Measure of Interclass Separability

In this paper we introduce the two-point correlation function as a measure of interclass separability. We present a theoretical study of this statistic in a general M-dimensional feature space and propose a fast algorithm for the efficient computation of it. We test the algorithm and illustrate the properties of the statistic using test data in 1D and 2D feature spaces and discuss the boundary effects of the feature space. We also present a discussion of the limitations of the proposed statistic and apply it to the assessment of inter-class separability in a texture segmentation context.