Search

The focus of recent few-shot learning research has been on the development of learning methods that can quickly adapt to unseen tasks with small amounts of data and low computational cost. In order to achieve higher performance in few-shot learning tasks, the generalizability of the method is essential to enable it generalize well from seen tasks to unseen tasks with limited number of samples. In this work, we investigate a new metric-based few-shot learning framework which transfers the knowledge from another effective classification model to produce well generalized embedding and improve the effectiveness in handling unseen tasks. The idea of our proposed Efficient-PrototypicalNet involves transfer learning, knowledge distillation, and few-shot learning. We employed a pre-trained model as a feature extractor to obtain useful features from tasks and decrease the task complexity. These features reduce the training difficulty in few-shot learning and increase the performance. Besides that, we further apply knowledge distillation to our framework and achieve extra performance improvement. The proposed Efficient-PrototypicalNet was evaluated on five benchmark datasets, i.e., Omniglot, miniImageNet, tieredImageNet, CIFAR-FS, and FC100. The proposed Efficient-PrototypicalNet achieved the state-of-the-art performance on most datasets in the 5-way K-shot image classification task, especially on the miniImageNet dataset.

Hand gesture provides a means for human to interact through a series of gestures. While hand gesture plays a significant role in human–computer interaction, it also breaks down the communication barrier and simplifies communication process between the general public and the hearing-impaired community. This paper outlines a convolutional neural network (CNN) integrated with spatial pyramid pooling (SPP), dubbed CNN–SPP, for vision-based hand gesture recognition. SPP is discerned mitigating the problem found in conventional pooling by having multi-level pooling stacked together to extend the features being fed into a fully connected layer. Provided with inputs of varying sizes, SPP also yields a fixed-length feature representation. Extensive experiments have been conducted to scrutinize the CNN–SPP performance on two well-known American sign language (ASL) datasets and one NUS hand gesture dataset. Our empirical results disclose that CNN–SPP prevails over other deep learning-driven instances.

Herbs are plants with savory or aromatic properties that are widely used for flavoring, food, medicine or perfume. The worldwide use of herbal products for healthcare has increased tremendously over the past decades. The plethora of herb species makes recognizing the herbs remains a challenge. This has spurred great interests among the researchers on pursuing artificial intelligent methods for herb classification. This paper presents a convolutional neural network (CNN) for herb classification. The proposed CNN consists of two convolution layers, two max pooling layers, a fully-connected layer and a softmax layer. The ReLU activation function and dropout regularization are leveraged to improve the performance of the proposed CNN. A dataset with 4067 herb images was collected for the evaluation purposes. The proposed CNN model achieves an accuracy of above 93% despite the fact that some herbs are visually similar.

Human activity recognition aims to determine the actions or behavior of a person based on the time series data. In recent year, more large human activity recognition datasets are available as it can be collected in easier and cheaper ways. In this work, a 1D Convolutional Neural Network with Long Short-Term Memory Network for human activity recognition is proposed. The 1D Convolutional Neural Network is employed to learn high-level representative features from the accelerometer and gyroscope signal data. The Long Short-Term Memory network is then used to encode the temporal dependencies of the features. The final classification is performed with a softmax classifier. The proposed 1D Convolutional Neural Network with Long Short-Term Memory Network is evaluated on MotionSense, UCI-HAR, and USC-HAD datasets. The class distributions of these datasets are imbalanced. In view of this, adjusted class weight is proposed to mitigate the imbalanced class issue. Furthermore, early stopping is utilized to reduce the overfitting in the training. The proposed method achieved promising performance on MotionSense, UCI-HAR, and USC-HAD datasets, with F1-score of 98.14%, 91.04%, and 76.42%, respectively.

Stock market prediction is a difficult task as it is extremely complex and volatile. Researchers are exploring methods to obtain good performance in stock market prediction. In this paper, we propose a Stacked Bidirectional Long Short-Term Memory (SBLSTM) network for stock market prediction. The proposed SBLSTM stacks three bidirectional LSTM networks to form a deep neural network model that can gain better prediction performance in the stock price forecasting. Unlike LSTM-based methods, the proposed SBLSTM uses bidirectional LSTM layers to obtain the temporal information in both forward and backward directions. In this way, the long-term dependencies from the past and future stock market values are encapsulated. The performance of the proposed SBLSTM is evaluated on six datasets collected from Yahoo Finance. Additionally, the proposed SBLSTM is compared with the state-of-the-art methods using root mean square error. The empirical studies on six datasets demonstrates that the proposed SBLSTM outperforms the state-of-the-art methods.

With the advancement in deep learning, high-resolution face recognition has achieved outstanding performance that makes it widely adopted in many real-world applications. Face recognition plays a vital role in visual surveillance systems. However, the images captured by the security cameras are at low resolution causing the performance of the low-resolution face recognition relatively inferior. In view of this, we propose an enhanced AlexNet with Super-Resolution and Data Augmentation (SRDA-AlexNet) for low-resolution face recognition. Firstly, image super-resolution improves the quality of the low-resolution images to high-resolution images. Subsequently, data augmentation is applied to generate variations of the images for larger data size. An enhanced AlexNet with batch normalization and dropout regularization is then used for feature extraction. The batch normalization aims to reduce the internal covariate shift by normalizing the input distributions of the mini-batches. Apart from that, the dropout regularization improves the generalization capability and alleviates the overfitting of the model. The extracted features are then classified using k-Nearest Neighbors method for low-resolution face recognition. Empirical results demonstrate that the proposed SRDA-AlexNet outshines the methods in comparison.

Computer vision has penetrated many domains, for instance, security, sports, health and medicine, agriculture, transportation, manufacturing, retail, and so like. One of the computer vision tasks is character recognition. In this work, a visually similar handwritten Chinese character dataset is collected. Subsequently, an enhanced convolutional neural network is proposed for the recognition of visually similar handwritten Chinese characters. The convolutional neural network is enhanced by the dropout regularization and early stopping mechanism to reduce the overfitting problem. The Adam optimizer is also leveraged to accelerate and optimize the training process of the convolutional neural network. The empirical results demonstrate that the enhanced convolutional neural network achieves a 97% accuracy, thus corroborate it has better discriminating power in visually similar handwritten Chinese character recognition.

Traffic sign recognition is a computer vision technique to recognize the traffic signs put on the road. In this paper, a traffic sign dataset with approximately 5000 images is collected. This paper presents an ablation analysis of Multilayer Perceptron and Convolutional Neural Networks in traffic sign recognition. The ablation analysis studies the effects of different architectures of Multilayer Perceptron and Convolutional Neural Networks, batch normalization, and dropout. A total of 8 different models are reviewed and their performance is studied. The experimental results demonstrate that Convolutional Neural Networks outperform Multilayer Perceptron in general. Leveraging dropout layer and batch normalization is effective in improving the stability of the model and achieved 98.62% accuracy in traffic sign recognition.

In the past decades, information and communication technology has developed rapidly. Therefore, social media has become the main platform for people to share and spread information to others. Although social media has brought a lot of convenience to people, fake news also spread more rapidly than before. This situation has brought a destructive impact to people. In view of this, we propose a hybrid model of Convolutional Neural Network and Long Short-Term Memory for fake news detection. The Convolutional Neural Network model plays the role of extracting representative high-level sequence features whereas the Long Short-Term Memory model encodes the long-term dependencies of the sequence features. Two regularization techniques are applied to reduce the model complexity and to mitigate the overfitting problem. The empirical results demonstrate that the proposed Convolutional Neural Network -Long Short-Term Memory model yields the highest F1-score on four fake news datasets.

Information and communication technology has evolved rapidly over the past decades, with a substantial development being the emergence of social media. It is the new norm that people share their information instantly and massively through social media platforms. The downside of this is that fake news also spread more rapidly and diffuse deeper than before. This has caused a devastating impact on people who are misled by fake news. In the interest of mitigating this problem, fake news detection is crucial to help people differentiate the authenticity of the news. In this research, an enhanced convolutional neural network (CNN) model, referred to as Fake News Net (FN-Net) is devised for fake news detection. The FN-Net consists of more pairs of convolution and max pooling layers to better encode the high-level features at different granularities. Besides that, two regularization techniques are incorporated into the FN-Net to address the overfitting problem. The gradient descent process of FN-Net is also accelerated by the Adam optimizer. The empirical studies on four datasets demonstrate that FN-Net outshines the original CNN model.

In recent years, facial emotion recognition (FER) has become a prevalent research topic as it can be applied in various areas. The existing FER approaches include handcrafted feature-based methods (HCF) and deep learning methods (DL). HCF methods rely on how good the manual feature extractor can perform. The manually extracted features may be exposed to bias as it depends on the researcher’s prior knowledge of the domain. In contrast, DL methods, especially Convolutional Neural Network (CNN), are good at performing image classification. The downfall of DL methods is that they require extensive data to train and perform recognition efficiently. Hence, we propose a deep learning method based on transfer learning of pre-trained AlexNet architecture for FER. We perform full model finetuning on the Alexnet, which was previously trained on the Imagenet dataset, using emotion datasets. The proposed model is trained and tested on two widely used facial expression datasets, namely extended Cohn-Kanade (CK+) dataset and FER dataset. The proposed framework outperforms the existing state-of-the-art methods in facial emotion recognition by achieving the accuracy of 99.44% and 70.52% for the CK+ dataset and the FER dataset.

This paper presents an isolated sign language recognition system that comprises of two main phases: hand tracking and hand representation. In the hand tracking phase, an annotated hand dataset is used to extract the hand patches to pre-train Convolutional Neural Network (CNN) hand models. The hand tracking is performed by the particle filter that combines hand motion and CNN pre-trained hand models into a joint likelihood observation model. The predicted hand position corresponds to the location of the particle with the highest joint likelihood. Based on the predicted hand position, a square hand region centered around the predicted position is segmented and serves as the input to the hand representation phase. In the hand representation phase, a compact hand representation is computed by averaging the segmented hand regions. The obtained hand representation is referred to as “Hand Energy Image (HEI)”. Quantitative and qualitative analysis show that the proposed hand tracking method is able to predict the hand positions that are closer to the ground truth. Similarly, the proposed HEI hand representation outperforms other methods in the isolated sign language recognition.

Stock market prediction has been a challenging task for machine due to time series analysis is needed. In recent years, deep neural networks have been widely applied in many financial time series tasks. Typically, deep neural networks require huge amount of data samples to train a good model. However, the data samples for stock market is limited which caused the networks prone to overfitting. In view of this, this paper leverages deep neural networks with ensemble learning to address this problem. We propose ensemble of Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), and 1DConvNet with LSTM (Conv1DLSTM) to predict the stock market price, named EnsembleDNNs. The performance of the proposed EnsembleDNNs is evaluated with stock market of several companies. The experiment results show encouraging performance as compared to other baselines.

Sound waves are a form of energy produced by a vibrating object that travels through the medium that can be heard. Generally, the sound is used in human communication, music, alert, and so on. Furthermore, it also helps us to understand what are the events that occurring in the moment, and thereby, provide us hints to understand what is happening around us. This has prompt researchers to study on how humans understand what event is occurring based on the sound waves. In recent years, researchers also study on how to equip the machine with this ability, i.e. acoustic event detection. This study focuses on the acoustic event detection which leverage both frequency spectrogram technique and deep learning methods. Initially, a spectrogram image is generated from the acoustic data by using the frequency spectrogram technique. Then, the generated frequency spectrogram is fed into a pre-trained MobileNet model to extract robust features representations. In this work, 1 Dimensional Convolutional Neural Network (1D-CNN) is adopted to train a model for acoustic event detection. The feature representations are extracted from a pre-trained MobileNet. The proposed 1D-CNN consist of several alternatives of convolution and pooling layers. The last pooling layer is flattened and fed into a fully connected layer to classify the events. Dropout is employed to prevent overfitting. The proposed frequency spectrogram with pre-trained MobileNet and 1D-CNN is then evaluated with three datasets, which are Soundscapes1, Soundscapes2, and UrbanSound8k. From the experimental results, the proposed method obtained 81, 86, and 70 F1-score, for Soundscapes1, Soundscapes2, and UrbanSound8k, respectively.

This paper presents a video-based human action recognition method leveraging deep learning model. Prior to the filtering phase, the input images are pre-processed by converting them into grayscale images. Thereafter, the region of interest that contains human performing action are cropped out by a pre-trained pedestrian detector. Next, the region of interest will be resized and passed as the input image to the filtering phase. In this phase, the filter kernels are trained using Sparse Autoencoder on the natural images. After obtaining the filter kernels, convolution operation is performed in the input image and the filter kernels. The filtered images are then passed to the feature extraction phase. The Histogram of Oriented Gradients descriptor is used to encode the local and global texture information of the filtered images. Lastly, in the classification phase, a Modified Hausdorff Distance is applied to classify the test sample to its nearest match based on the histograms. The performance of the deep learning algorithm is evaluated on three benchmark datasets, namely Weizmann Action Dataset, CAD-60 Dataset and Multimedia University (MMU) Human Action Dataset. The experimental results show that the proposed deep learning algorithm outperforms other methods on the Weizmann Dataset, CAD-60 Dataset and MMU Human Action Dataset with recognition rates of 100%, 88.24% and 99.5% respectively.

Object recognition has spurred much attention in recent years. The fact that computers are now able to detect and recognize objects has made Artificial Intelligence field, especially machine learning grow very rapidly. The proposed framework uses Deep Convolutional Neural Network (DCNN) that is based on ResNet 50 architecture. Due to the limited computational resources to train the whole model, the ResNet model is imitated and the pre-trained weights are imported. Thereafter, the last few layers of the model are trained on three datasets that have been acquired online. This process is called fine-tuning a pre-trained model. It is one of the most common approaches in building a DCNN architecture. The dataset that was used to evaluate the performance of the model are ETHZ-FOOD101, UECFOOD100 and UECFOOD256. The parameter setting and results of the proposed method are also presented in this paper.

In this paper, we propose a hand tracking method which was inspired by the notion of the four dukkha: birth, aging, sickness and death (BASD) in Buddhism. Based on this philosophy, we formalize the hand tracking problem in the BASD framework, and apply it to hand track hand gestures in isolated sign language videos. The proposed BASD method is a novel nature-inspired computational intelligence method which is able to handle complex real-world tracking problem. The proposed BASD framework operates in a manner similar to a standard state-space model, but maintains multiple hypotheses and integrates hypothesis update and propagation mechanisms that resemble the effect of BASD. The survival of the hypothesis relies upon the strength, aging and sickness of existing hypotheses, and new hypotheses are birthed by the fittest pairs of parent hypotheses. These properties resolve the sample impoverishment problem of the particle filter. The estimated hand trajectories show promising results for the American sign language.

Hand gesture recognition (HGR) serves as a fundamental way of communication and interaction for human being. While HGR can be applied in human computer interaction (HCI) to facilitate user interaction, it can also be utilized for bridging the language barrier. For instance, HGR can be utilized to recognize sign language, which is a visual language represented by hand gestures and used by the deaf and mute all over the world as a primary way of communication. Hand-crafted approach for vision-based HGR typically involves multiple stages of specialized processing, such as hand-crafted feature extraction methods, which are usually designed to deal with particular challenges specifically. Hence, the effectiveness of the system and its ability to deal with varied challenges across multiple datasets are heavily reliant on the methods being utilized. In contrast, deep learning approach such as convolutional neural network (CNN), adapts to varied challenges via supervised learning. However, attaining satisfactory generalization on unseen data is not only dependent on the architecture of the CNN, but also dependent on the quantity and variety of the training data. Therefore, a customized network architecture dubbed as enhanced densely connected convolutional neural network (EDenseNet) is proposed for vision-based hand gesture recognition. The modified transition layer in EDenseNet further strengthens feature propagation, by utilizing bottleneck layer to propagate the features being reused to all the feature maps in a bottleneck manner, and the following Conv layer smooths out the unwanted features. Differences between EDenseNet and DenseNet are discerned, and its performance gains are scrutinized in the ablation study. Furthermore, numerous data augmentation techniques are utilized to attenuate the effect of data scarcity, by increasing the quantity of training data, and enriching its variety to further improve generalization. Experiments have been carried out on multiple datasets, namely one NUS hand gesture dataset and two American Sign Language (ASL) datasets. The proposed EDenseNet obtains 98.50% average accuracy without augmented data, and 99.64% average accuracy with augmented data, outperforming other deep learning driven instances in both settings, with and without augmented data.

Gait has unique advantage at a distance when other biometrics cannot be used since they are at too low resolution or obscured, as commonly observed in visual surveillance systems. This paper provides a survey of the technical advancements in vision-based gait recognition. A wide range of publications are discussed in this survey embracing different perspectives of the research in this area, including gait feature extraction, classification schemes and standard gait databases. There are two major groups of the state-of-the-art techniques in characterizing gait: Model-based and motion-free. The model-based approach obtains a set of body or motion parameters via human body or motion modeling. The model-free approach, on the other hand, derives a description of the motion without assuming any model. Each major category is further organized into several subcategories based on the nature of gait representation. In addition, some widely used classification schemes and benchmark databases for evaluating performance are also discussed.

A normalized histogram of optical flow as a hand representation of the sign language.Block-based histogram provides spatial information and local translation invariant.Block-based histogram of optical flow enables sign language length invariance. In this paper, we propose a block-based histogram of optical flow (BHOF) to generate hand representation in sign language recognition. Optical flow of the sign language video is computed in a region centered around the location of the detected hand position. The hand patches of optical flow are segmented into M spatial blocks, where each block is a cuboid of a segment of a frame across the entire sign gesture video. The histogram of each block is then computed and normalized by its sum. The feature vector of all blocks are then concatenated as the BHOF sign gesture representation. The proposed method provides a compact scale-invariant representation of the sign language. Furthermore, block-based histogram encodes spatial information and provides local translation invariance in the extracted optical flow. Additionally, the proposed BHOF also introduces sign language length invariancy into its representation, and thereby, produce promising recognition rate in signer independent problems.

A fusion of median and mode filtering for better background model.A serial particle filter that can better detect and track the object of interest.A novel covariance matrix feature for isolated sign language representation. As is widely recognized, sign language recognition is a very challenging visual recognition problem. In this paper, we propose a feature covariance matrix based serial particle filter for isolated sign language recognition. At the preprocessing stage, the fusion of the median and mode filters is employed to extract the foreground and thereby enhances hand detection. We propose to serially track the hands of the signer, as opposed to tracking both hands at the same time, to reduce the misdirection of target objects. Subsequently, the region around the tracked hands is extracted to generate the feature covariance matrix as a compact representation of the tracked hand gesture, and thereby reduce the dimensionality of the features. In addition, the proposed feature covariance matrix is able to adapt to new signs due to its ability to integrate multiple correlated features in a natural way, without any retraining process. The experimental results show that the hand trajectories as obtained through the proposed serial hand tracking are closer to the ground truth. The sign gesture recognition based on the proposed methods yields a 87.33% recognition rate for the American Sign Language. The proposed hand tracking and feature extraction methodology is an important milestone in the development of expert systems designed for sign language recognition, such as automated sign language translation systems.

In this paper, we present a gait recognition method using convolutional features and histograms of temporal gradients. The method comprises three stages, namely the convolutional stage, temporal gradient stage and classification stage. In the convolutional stage, the video frames are convolved with a set of pre-learned filters. This is followed by the temporal gradient stage. In this stage, the gradient of each convolved frame in time axis is calculated. Unlike histograms of oriented gradients that accumulate the gradients in the spatial domain, the proposed histogram of temporal gradients encodes the gradients in the spatial and temporal domain. The histogram of temporal gradients captures the gradient patterns of every pixel over the temporal axis throughout the video sequence. By doing so, the feature encodes both spatial and temporal information in the gait cycle. Finally, in the classification stage, a majority voting classification with Euclidean distance is performed for gait recognition. Experimental results show that the proposed method outperforms the state-of-the-art methods.

Visual representation is crucial for a visual tracking method's performances. Conventionally, visual representations adopted in visual tracking rely on hand-crafted computer vision descriptors. These descriptors were developed generically without considering tracking-specific information. In this paper, we propose to learn complex-valued invariant representations from tracked sequential image patches, via strong temporal slowness constraint and stacked convolutional autoencoders. The deep slow local representations are learned offline on unlabeled data and transferred to the observational model of our proposed tracker. The proposed observational model retains old training samples to alleviate drift, and collect negative samples which are coherent with target's motion pattern for better discriminative tracking. With the learned representation and online training samples, a logistic regression classifier is adopted to distinguish target from background, and retrained online to adapt to appearance changes. Subsequently, the observational model is integrated into a particle filter framework to peform visual tracking. Experimental results on various challenging benchmark sequences demonstrate that the proposed tracker performs favourably against several state-of-the-art trackers., Pattern Recognition (Elsevier), 2015

This paper presents a gait recognition method using the combination of motion history image (MHI), binarized statistical image features (BSIF) and histograms of oriented gradients (HOG). The method first encodes the motion pattern and direction of the gait cycle in motion history image. Subsequently, performing convolution on the motion history image using pre-learnt filters as kernel, binarized statistical image features are generated by summing the convolution output images. Histograms of oriented gradients are then computed on binarized statistical image features. Gait signature of a gait cycle is attained by accumulating all the HOG descriptors. Experimental result shows that the proposed method performs promisingly in gait recognition.

Finger spelling is a way of communication by expressing words using hand signs in order to ensure deaf and dumb community can communicate with others effectively. Therefore, a system that can understand finger spelling is needed. As a result of that, this work is conducted to primarily develop a tutoring system for finger spelling. To develop a robust real-time finger spelling tutoring system, it is necessary to ensure the accuracy of the finger spelling recognition. Even though there are existing solutions available for a decade, but most of them are just focusing on improving accuracy rate without implementing their solutions as a complete tutoring system for finger spelling. Consequently, it inspires this research project to develop a tutoring system for finger spelling. Microsoft Kinect sensor is used to acquire color images and depth images of the finger spells. Depth images are used to perform segmentation on the color images. After that, the segmented images are used as input and pass into a two hidden layers backpropagation neural network for classification.

This paper proposes a Chinese character flashcard recognizer for kids learning purposes. Firstly, the system will capture an image of the flashcard shown to the webcam. Secondly, a preprocessing procedure based on thresholding is performed to remove noisy edges and convert the color image into binary image. Subsequently, a Principal Component Analysis based feature extraction is conducted to encode the image into a compact representation. Lastly, character recognition is performed using Euclidean distance. The kids are able to learn Chinese characters with their Pinyin and meaning using the system. From the experiments, promising recognition results are achieved.

A fuzzy model based on enhanced supervised fuzzy clustering algorithm is presented in this paper. Supervised fuzzy clustering algorithm by Janos Abonyi and Ferenc Szeifert in the year 2003 allows each rule to represent more than one output with different probabilities for each output. This algorithm implements k-means to initialize the fuzzy model. The main drawbacks of this approach are the number of clusters is unknown and the initial positions of clusters are randomly generated. In this work, the initialization is performed by global k-means algorithm [1] which can autonomously determine the actual number of clusters needed and give deterministic clustering result. In addition, fast global k-means [1] is presented to improve the computation time. Besides that, when collecting input data in a feature vector way, it might occur that some of the feature values are lost for a particular vector due to a faulty reading sensor. To deal with missing values in enhanced supervised fuzzy clustering, the efficient way is imputation during data preprocessing. The modified of optimal completion strategy is presented to solve this problem. This method allows imputation of missing data with high reliability and accuracy. The autonomous and deterministic enhanced supervised fuzzy clustering using supervised Gath-Geva clustering method and the modified of optimal completion strategy can be derived from the unsupervised Gath-Geva algorithm. The proposed algorithm is successfully justified based on benchmark data sets and a real vibration data which was collected from U.S. Navy CH-46E helicopter aft gearbox called Westland.

Most of the current automated essay scoring (AES) systems are trained using manually graded essays from a specific prompt. These systems experience a drop in accuracy when used to grade an essay from a different prompt. Obtaining a large number of manually graded essays each time a new prompt is introduced is costly and not viable. We propose domain adaptation as a solution to adapt an AES system from an initial prompt to a new prompt. We also propose a novel domain adaptation technique that uses Bayesian linear ridge regression. We evaluate our domain adaptation technique on the publicly available Automated Student Assessment Prize (ASAP) dataset and show that our proposed technique is a competitive default domain adaptation algorithm for the AES task.

We propose a two-phase framework to adapt existing relation extraction classifiers to extract relations for new target domains. We address two challenges: negative transfer when knowledge in source domains is used without considering the differences in relation distributions; and lack of adequate labeled samples for rarer relations in the new domain, due to a small labeled data set and imbalance relation distributions. Our framework leverages on both labeled and unlabeled data in the target domain. First, we determine the relevance of each source domain to the target domain for each relation type, using the consistency between the clustering given by the target domain labels and the clustering given by the predictors trained for the source domain. To overcome the lack of labeled samples for rarer relations, these clusterings operate on both the labeled and unlabeled data in the target domain. Second, we trade-off between using relevance-weighted sourcedomain predictors and the labeled target data. Again, to overcome the imbalance distribution, the source-domain predictors operate on the unlabeled target data. Our method outperforms numerous baselines and a weakly-supervised relation extraction method on ACE 2004 and YAGO.

There are lots of ways to perform object recognition. This paper is part of a project studying object recognition. The project is intended as a starting point to further learning about object recognition. Therefore, moment invariants are studied as a good starting point. Hu moment invariant methods and Zernike moment invariant methods are implemented and compared. Zernike moment invariants are shown to outperform Hu moment invariants.

Multimodal biometrics are always adopted to improve the recognition performance of single modality biometric systems. Besides introducing more discriminating power to the biometric system, integrating multiple modalities also leads to the curse of dimensionality problem. In this paper, we engage the minimal redundancy maximal relevance criterion to reduce the dimensionality of the feature vector. The minimal redundancy maximal relevance criterion is a feature selection criterion that aims to retain the most relevant elements while discarding the other redundant elements. Our experiments show that, with only 15% of the original feature length, minimal redundancy maximal relevance criterion-based features are able to perform similarly well or even better than the baseline results.

As visual surveillance systems gain wider usage in a variety of fields, it is important that they are capable of interpreting scenes automatically, also known as “human motion analysis” (HMA). However, existing HMA methods are too domain specific and computationally expensive. This paper proposes a general purpose HMA method that is based on the idea that human beings tend to exhibit erratic motion patterns during abnormal situations. Limb movements are characterized using the statistics of angular and linear displacements. In addition, the method is enhanced via the use of the entropy of the Fourier spectrum to measure the randomness of subject’s motions. Various experiments have been conducted and the results indicate that the proposed method has very high classification accuracy in identifying anomalous behavior.

As visual surveillance systems are gaining wider usage in a variety of fields, they need to be embedded with the capability to interpret scenes automatically, which is known as human motion analysis (HMA). However, existing HMA methods are too domain specific and computationally expensive. This paper proposes a general purpose HMA method. It is based on the idea that human beings tend to exhibit random motion patterns during abnormal situations. Hence, angular and linear displacements of limb movements are characterized using basic statistical quantities. In addition, it is enhanced with the entropy of the Fourier spectrum to measure the randomness of the abnormal behavior. Various experiments have been conducted and prove that the proposed method has very high classification accuracy in identifying anomalous behavior.

As visual surveillance systems are gaining wider usage in a variety of fields, they need to be embedded with the capability to interpret scenes automatically, which is known as human motion analysis (HMA). However, existing HMA methods are too domain specific and computationally expensive. This paper proposes a general purpose HMA method. It is based on the idea that human beings tend to exhibit random motion patterns during abnormal situations. Hence, angular and linear displacements of limb movements are characterized using basic statistical quantities. In addition, it is enhanced with the entropy of the Fourier spectrum to measure the randomness of the abnormal behavior. Various experiments have been conducted and prove that the proposed method has very high classification accuracy in identifying anomalous behavior.

This paper explores the use of Bayesian online classifiers to classify text documents. Empirical results indicate that these classifiers are comparable with the best text classification systems. Furthermore, the online approach offers the advantage of continuous learning in the batch-adaptive text filtering task.

People work with computers almost anytime, everywhere in the current trend. However, continuously controlling a computer with mouse for a long time might cause much strains to people’s wrist. This work proposes a touchless finger mouse using webcam. A marker with different colours representing different actions is used. The webcam will capture the information on the marker and trigger the associated actions. This prototype is proven to be able to perform most of the actions a normal mouser can perform.