Deep learning has emerged as a pivotal technology across various domains, demonstrating remarkable performance. However, its susceptibility to security threats, particularly adversarial samples, poses a significant concern. These samples can manipulate inputs slightly, deceiving models such as those used in image or speech recognition and potentially leading to incorrect predictions. This undermines the reliability of deep learning in critical applications. In this paper, we propose an effective method utilizing Autoencoder to detect and intercept audio adversarial attacks before they are input to speech recognition models. The proposed approach first uses clean audio data to train the Autoencoder model, then isolates adversarial samples from clean ones by comparing their features against normal features encoded in the Autoencoder. Our method does not require prior knowledge about the target automatic speech recognition model or attack methods. Experimental results show that it can effectively discriminate adversarial attack samples from clean ones with high accuracy.

The selection of violin fingerings is influenced by factors such as musical context, skill level, and personal taste. Current deep-learningbased models, relying solely on symbolic data, are able to generate playable fingerings but struggle to capture the personal nuances of musical performance, which only lie in the audio data. To address this limitation, we introduce a novel model that incorporates both audio and symbolic data, allowing users to upload music scores and their corresponding violinist recordings to obtain personalized fingerings related to the audio data. To simulate such a real-world application scenario, we also collect a new dataset from online audios. The experiment results demonstrate the superiority of our proposed method over previous symbolic-based methods, even in the situations involving multiple instruments in audio.

Existing Siamese network-based trackers rely on stable appearance features extracted from the target object. However, such features might not be available during tracking due to non-digit appearance deformation and severe occlusion, which result in drift problems. In this paper, we propose a background-augmented tracking network that incorporates background information surrounding the target to make up for missing or deformed target features during the matching process. A novel Background Relation Network (BRNet) is designed to effectively encode and match the background information surrounding candidate objects in the search region to help identify the correct target, and thus avoid tracking error. BRNet can complement the base tracker when reliable target features cannot be obtained. Experiments on the OTB, VOT, and UAV123 datasets demonstrate that the proposed method achieves superior performance over existing state-of-the-art methods while maintaining reasonable real-time speed.

Background subtraction is an example of a moving object detection technique that uses machine vision systems. Conventional moving object detection methods need complicated thresholds for background modeling to address changes in illumination. This paper proposes a novel background modeling approach without thresholding based on a bit-planes method, which fully utilizes color characteristics through spatial and temporal-based improvement. The proposed idea is effective and efficiently solving for shadow disturbance and brightness changes. We evaluate our proposed method using several challenging indoor and outdoor sequences from the CDNET 2014 dataset. The experiments show that the proposed idea typically achieves a higher rate of detection accuracy than those of the current state-of-the-art approaches.

Indocyanine green (ICG) angiography is an imaging method for doctors to observe choroidal abnormalities in human eyes. The ICG angiograms typically exhibit inhomogeneous illumination, which poses serious difficulties for the development of computer-aided diagnostic tools. In this paper, we propose a novel illumination normalization method to alleviate the inhomogeneous illumination in ICG video angiograms. In particular, we first align the viewpoint of the input ICG video angiogram using an image registration method. Then, we acquire temporal information using time-dependent intrinsic image and compute the corresponding illumination image. Finally, we correct inhomogeneous illumination from the illumination image by estimating contrast and luminosity distortion. We have conducted extensive evaluation using ICG video angiograms of 60 patients. Two video quality assessment methods are utilized to evaluate the performance of our proposed illumination normalization method. The results show that our proposed method can help improve the visual quality of ICG video angiogram. Visual evaluation by a human expert also confirms that our method yields better illumination normalization results.

In ophthalmology, aligning images in indocyanine green and fluorescein angiograph sequences is important for the treatment of subretinal lesions. This paper introduces an algorithm that is tailored to align jointly in a common reference space all the images in an angiogram sequence containing both modalities. To overcome the issues of low image contrast and low signal-to-noise ratio for late-phase images, the structural similarity between two images is enhanced using Gabor wavelet transform. Image pairs are pairwise registered and the transformations are simultaneously and globally adjusted for a mutually consistent joint alignment. The joint registration process is incremental and the success depends on the correctness of matches from the pairwise registration. To safeguard the joint process, our system performs the consistency test to exclude incorrect pairwise results automatically to ensure correct matches as more images are jointly aligned. Our dataset consists of 60 sequences of polypoidal choroidal vasculopathy collected by the EVEREST Study Group. On average, each sequence contains 20 images. Our algorithm successfully pairwise registered 95.04% of all image pairs, and joint registered 98.7% of all images, with an average alignment error of 1.58 pixels.

In this paper, we propose a novel method for extracting brain region from T1-weighted magnetic resonance imaging. Our proposed method is based on finding a seed point which is located on brain tissue and then perform region growing. In particular, we firstly analyze intensity distributions of different cerebral tissues (including brain, cerebrospinal fluid, scalp and marrow) at middle height. We also measure the distance from image center to brain edge at the same height of brain scan. In the experimental validation, we have conducted experiment on the OASIS dataset and perform comparison with the state-of-the-art methods. Our propose method achieves the best performance on the OASIS dataset.

For financial institutions, the ability to predict or forecast business failures is crucial, as incorrect decisions can have direct financial consequences. Bankruptcy prediction and credit scoring are the two major research problems in the accounting and finance domain. In the literature, a number of models have been developed to predict whether borrowers are in danger of bankruptcy and whether they should be considered a good or bad credit risk. Since the 1990s, machine-learning techniques, such as neural networks and decision trees, have been studied extensively as tools for bankruptcy prediction and credit score modeling. This paper reviews 130 related journal papers from the period between 1995 and 2010, focusing on the development of state-of-the-art machine-learning techniques, including hybrid and ensemble classifiers. Related studies are compared in terms of classifier design, datasets, baselines, and other experimental factors. This paper presents the current achievements and limitations associated with the development of bankruptcy-prediction and credit-scoring models employing machine learning. We also provide suggestions for future research.

This paper presents a method for recognizing sign language using hand movement trajectory. By applying Kernel Principal Component Analysis (KPCA), the motion trajectory data are firstly mapped into a higher-dimensional feature space for analysis. The advantage of using high dimensionality is that it allows a more flexible decision boundary and thus helps us to achieve better classification accuracy. Then we perform the Nonparametric Discriminant Analysis (NDA) in the higher-dimensional feature space, so that the most helpful information can be extracted. We have tested the proposed method using the Australian Sign Language (ASL) data set. The results demonstrate that our approach outperforms the current state-of-the-art for trajectory-based sign language recognition.

In this paper, we propose a novel framework for video-based human action recognition, which can effectively resolve the difficulty caused by large variations within each action category. We first use the cloud of interest points to represent human action, due to its effectiveness in extracting spatio-temporal information necessary to reliability distinguish each action. Then, we perform an efficient local learning on the extracted features to learn locally adaptive classifiers. Specifically, a local classifier is specifically trained for each training sample. A local classifier could better describe the local data distribution and thus adopting multiple local classifiers would lead to better classification accuracy. We conduct several experiments on the KTH dataset and obtain very inspiring results. In particular, our approach achieves comparable performance to that of the state-of-the-art methods. Compared with a global learning method, i.e., the AdaBoost, the local learning provides significantly better accuracy with little additional cost in training time.

This paper presents a novel scheme for human action recognition. First of all, we employ the curvature estimation to analyze human posture patterns and to yield the discriminative feature sequences. The feature sequences are further represented into sets of strings. Consequently, we can solve human action recognition problem by the string matching technique. In order to boost the performance of string matching, we apply the Nonparametric Weighted Feature Extraction (NWFE) to compact the string representation. Finally, we train a Bayes classifier to perform action recognition. Unlike traditional approaches using the nearest neighbor rule, our proposed scheme can classify the human actions more efficiently while maintaining high accuracy. The experiment results show that the proposed scheme is efficient and accurate in human action recognition.

In this rapid development of advanced computer technology era, the trend for interaction with the input devices has shifted toward the incorporation of intelligent technology gradually. To facilitate the intuitive paradigm for human to use digital equipment, the traditional input devices such as keyboards and mice are augmented or even replaced by various novel mechanisms. The provision of the intuitive functions by using a variety of end devices such as camera, RFID, WiiMote enhances the interaction and playfulness of information presentation. This paper presents a set of tools to help developers create interactive presentations with machine intelligence. The design rationale of the set of tools called the “U-Garden” was described. We also illustrate how the U-Garden could be adopted to synchronize the hardware devices and media content to build an intelligent interaction. To test the performance of the U-Garden, the tools were applied to interactive presenting applications to demonstrate the content presenting power for digital archives. Finally, experiments with user experience on our proposed content presenting system were conducted and discussed. The results reveal positive responses that users are able to appreciate this intelligent presenting of the content of digital archives through our tools.

This paper addresses an old, yet challenging issue - curvature estimation from discrete sampling points over a curve. We introduce a novel algorithm based on performing line integrals. The proposed method is computationally more efficient than the previous integration-based methods because of the constant computation time. Qualitative tests on synthesized shapes in the presence of noise are performed, which show the robustness of our approach. Also, the discriminative capability of the estimated curvature is evaluated by conducting experiments on the FRGC (Face Recognition Grand Challenge) v2.0 dataset which contains 4007 3D facial images recorded from 466 subjects. The results show that recognition performance is significantly improved by using our curvature estimation method. This novel approach presents potential for a broad class of multimedia applications.

In this paper, we present a method to automatically extract the vessel segments and construct the vascular tree with anatomical realism from a color retinal image. The significance of the work is to assist in clinical studies of diagnosis of cardio-vascular diseases, such as hypertension,which manifest abnormalities in either venous and/or arterial vascular systems. To maximize the completeness of vessel extraction, we introduce vessel connectiveness measure to improve on an existing algorithm which applies multiscale matched filtering and vessel likelihood measure.Vessel segments are grouped using extended Kalman filter to take into consideration continuities in curvature, width,and color changes at the bifurcation or crossover point. The algorithm is tested on five images from the DRIVE database,a mixture of normal and pathological images, and the results are compared with the ground truth images provided by a physician. The preliminary results show that our method reaches an average success rate of 92.1%.

Applications of summation invariant features to multi-region face recognition are explored in this work. Earlier, we have demonstrated the potential benefits of this approach. In this paper, we provide a systematic, thorough comparison of all the summation invariant features derived to-date, and propose a new multi-feature fusion approach to further improve the overall performance. We also identify summation invariant features that yield superior performance for face recognition applications. Special attention is given to the implementation of 3D summation invariants. Extensive experimental results with the FRGC (Face Recognition Grand Challenge) 2.0 data set confirms the advantage of summation invariant features for 3D face recognition.

Recently a novel family of geometrically invariant features, called summation invariant, has been developed and applied to object recognition. The range of this family of features is expanded here beyond the Euclidean and afline transformation groups to planar projective transformations. Whereas other methods require small changes in view, or collinear points, this method removes those limitations and allows recognition of general planar objects over wide ranges of viewpoint. The derivation of these new features requires the innovation of deriving the invariants in the homogeneous coordinate space, yet yields results formulated in terms of Cartesian coordinates. Simulations demonstrate the effectiveness of this new approach to object recognition under projective transformations like those encountered in camera networks.

The proposed system aims at providing a novel framework of speaker authentication using lip-motion data. The system is divided into three steps: feature extraction and modeling, model synthesis, and probabilistic model matching. First, the visual features are obtained by locating a region-of-interest and extracting geometric and textural features from the region. The extracted feature vectors are then classified by K-means clustering to obtain a reduced number of observation vectors. These observation vectors are fed into a set of Hidden Markov Models (HMMs) classifiers to capture the temporal characteristics of the features. The main contribution of this paper lies in the introduction of using random passwords in performing speaker verification. Our results have demonstrated that random passwords provide useful information for speaker verification. Also, by using the proposed method, we observe a significant improvement in verification rate.

We present a novel 3D face recognition method that incorporates summation invariant features extracted from multiple sub-regions of a facial range images, and optimal fusion of similarity scores between corresponding sub-regions. The key innovation of this paper is the development of the fusion-based face recognition algorithm that delivers significant performance enhancement while requiring very little computation. Experiments on the FRGC (Face Recognition Grand Challenge) version 2 dataset show that our algorithm improves the recognition performance significantly in the presence of facial expressions.

In this paper, we describe a fusion-based face recognition method that is able to compensate for facial expressions even when training samples contain only neutral expression. The similarity metric between two facial images are calculated by combining the similarity scores of the corresponding facial regions, e.g. the similarity between two mouths, the similarity between two noses, etc. In contrast with other approaches where equal weights are assigned on each region, a novel fusion method based on linear discriminant analysis (LDA) is developed to maximize the verification performance. We also conduct a comparative study on various face recognition schemes, including the FRGC baseline algorithm, the fusion of multiple regions by sum rule, and the fusion of multiple regions by LDA. Experiments on the FRGC (Face Recognition Grand Challenge) V2.0 dataset, containing 4007 face images recorded from 266 subjects, show that the proposed method significantly improves the verification performance in the presence of facial expressions.

This paper presents a novel 3-D multiregion face recognition algorithm that consists of new geometric summation invariant features and an optimal linear feature fusion method. A summation invariant, which captures local characteristics of a facial surface, is extracted from multiple subregions of a 3-D range image as the discriminative features. Similarity scores between two range images are calculated from the selected subregions. A novel fusion method that is based on a linear discriminant analysis is developed to maximize the verification rate by a weighted combination of these similarity scores. Experiments on the Face Recognition Grand Challenge V2.0 dataset show that this new algorithm improves the recognition performance significantly in the presence of facial expressions.

A novel family of 2D and 3D geometrically invariant features, called summation invariants is proposed for the recognition of the 3D surface of human faces. Focusing on a rectangular region surrounding the nose of a 3D facial depth map, a subset of the so called semi-local summation invariant features is extracted. Then the similarity between a pair of 3D facial depth maps is computed to determine whether they belong to the same person. Out of many possible combinations of these set of features, we select, through careful experimentation, a subset of features that yields best combined performance. Tested with the 3D facial data from the on-going Face Recognition Grand Challenge v1.0 dataset, the proposed new features exhibit significant performance improvement over the baseline algorithm distributed with the datase

A novel family of geometrically invariant features, called summation invariants are proposed for the recognition of the 3D surface of human faces. In particular, a 2D semi-local summation invariant feature is extracted from each column and each row of a rectangular region surrounding the nose of a 3D facial depth map. Through extensive experimentation, we empirically identify the most efficient 2D summation invariant features. We also investigate the proper pre-processing method for the 2D summation invariant features. Tested with the 3D facial data from the Face Recognition Grand Challenge v1.0 dataset, the proposed new features exhibit significant performance improvement over the baseline algorithm distributed with the dataset.

In this paper, we developed a family of 2D and 3D invariant features with applications to 3D human faces recognition. The main contributions of this paper are: (a) systematically deriving a family of novel features, called summation invariant that are invariant to Euclidean transformation in both 2D and 3D; (b) developing an effective method to apply summation invariant to the 3D face recognition problem. Tested with the 3D data from the face recognition grand challenge v1.0 dataset, the proposed new features exhibit achieves a performance that rivals the best 3D face recognition algorithms reported so far

A novel summation invariant of curves under transformation group action is proposed. This new invariant is less sensitive to noise than the differential invariant and does not require an analytical expression for the curve as the integral invariant does. We exploit this summation invariant to define a shape descriptor called a semi-local summation invariant and use it as a new feature for shape recognition. Tested on a database of noisy shapes of fish, it was observed that the summation invariant feature exhibited superior discriminating power compared to that of wavelet-based invariant features.

A novel summation invariant feature under transformation group action for 3D surface recognition is proposed, and its application to 3D face recognition is investigated. Based on a systematic mathematical procedure called moving frame, we derived the summation invariant feature that is invariant under affine transformation. Compared with classical differential invariants, such as the mean curvature or the Gaussian curvature, summation invariant feature is far less sensitive to observation noise in the data. A further enhancement leads to a new type of invariant 3D surface shape descriptor called a semi-local summation invariant. We demonstrate one important, potential application of this new feature to 3D human face recognition