Parinya Sanguansat

TF-RNF: A novel term weighting scheme for sports video classification

Determination of content importance is very important in achieving high quality classification. Term weighting schemes in text classification will be applied to classify videos by measuring importance of video contents. In other words, a video sequence can be treated as a document, and frames of a video are considered as words or terms which identify contents of a video. And to enhance the efficiency of video classification, this paper proposes a novel term weighting scheme, called the Term Frequency - Relevance and Non-relevance Frequency (TF-RNF) weighting. This technique can filter both relevant and non-relevant contents so as to reduce classification errors. Empirical evaluations of results show that the proposed technique significantly outperforms traditional techniques in sports video classification.

A Music Information System Based on Improved Melody Contour Extraction

In this paper, we propose a new melody contour extraction technique to improve Query-by-Humming. A critical issue of humming sound is noise interference from both environment and acquisition instrument. Furthermore most users are not professional singers so they cause the other query problems about variation of pitch and timing. Advantage of a proposed technique can reduce noise whereas makes pitch smoothing. Moreover, this method is used to enhance the query accuracy. Our technique consists of three steps as follows: Firstly, the melody contour is extracted from humming sound by Subharmonic-to-Harmonic Ratio (SHR). Subsequently, the melody contour is filtered and smoothed by statistical approach. Finally, Dynamic Time Warping (DTW) is applied to melody contour, for similarity measurement between humming sound and melody sequence. We use a new method to extract melody contour from pitch. Experimental results show that our proposed technique can perform better than traditional method.

Two-Dimensional Random Projection for Face Recognition

In this paper, the two-dimensional random projection (2DRP) is proposed to directly project the image matrix from high-dimensional space to low-dimensional space for recognition task. In traditional random projection framework, the projection matrix does not depend on the training data hence it can avoid the principal classification problems such as over-fitting, Small Sample Size (SSS), and singularity problems. However, the face images must be transformed to vectors before projection. In this way, the size of projection matrix will depend on the product of width and height of the image, that is very large and consumes lots of memory and computation time to process. Instead of the traditional projection, our method uses unilateral and bilateral projection for an image matrix directly, which applies the left and right projections to each image matrix side by side or simultaneously. Thus, the size of left and right projections will depend on only the height or width of an image, respectively. The memory and computation time of this method will be substantially reduced. After projection, we investigate the results of 2DRP by the nearest neighbor classifier. Our experiments on well-known face databases demonstrate the significant of our proposed method.

Developing an effective Thai Document Categorization Framework base on term relevance frequency weighting

Text Categorization is the process of automatically assigning predefined categories to free text documents. Feature weighting, which calculates feature (term) values in documents, is an important preprocessing technique in text categorization. In this paper, we purpose Thai Document Categorization Framework focusing on the comparison of various term weighting schemes, including Boolean, tf, tf-idf, tfc, ltc entropy and tf-rf weighting. We have evaluated these methods on Thai news article corpus with three supervised learning classifiers. We found tf-rf weighting most effective in our experiments with SVM NB and DT algorithms. Based on our experiments, using tf-rf weighting with SVM algorithm yielded the best performance with the F-measure equaling 95.9%.

Classification via k-means clustering and distance-based outlier detection

We propose a two-phase classification method. Specifically, in the first phase, a set of patterns (data) are clustered by the k-means algorithm. In the second phase, outliers are constructed by a distance-based technique and a class label is assigned to each pattern. The Knowledge Discovery Databases (KDD) Cup 1999 data set, which has been utilized extensively for development of intrusion detection systems, is used in our experiment. The results show that the proposed method is effective in intrusion detection.

A Novel Method for Query-by-Humming Using Distance Space

Signal of humming sound is the input which is important for the Query-by-Humming system. This input signal which has variable dimension depend on humming time interval will always affect the feature vector. It cannot be used with some classifiers, which require non-variable dimension of feature vector, such as Artificial Neural Network (ANN) or Support Vector Machine (SVM). Especially, SVM is good classifier and it might be appropriate for our work. Because of each signal of humming sound has variable dimension and length, this is the main problem which we would like to come up with the idea to solve it. We have an idea to create a new feature space that has the same dimension in order to use with SVM classifier. In this paper, we propose indirect feature, it is used distance between template and observation sequence for creating new feature vector. This technique can be briefly described: Firstly, templates are distributed in original feature space. When the observation sequence gets into this space, Dynamic Time Warping (DTW) will measure the distance between observation sequence and existing templates. These distance are used to get the new feature vector in new space, called distance space. In this way, all feature vectors are non-variable dimension therefore we used SVM and ANN classifier. The experimental results show that the new feature vector which is used by SVM classifier gives better results than ANN.

A Novel Approach for Measuring Video Similarity without Threshold and Its Application in Sports Video Categorization

One of the most important issues for measuring video similarity is the difficulty in identifying the optimal frame similarity threshold, which often tends to vary in an unpredictable pattern, and has to be manually determined. Moreover, most video data are huge files, which vary in terms of length and amount of data, resulting in time-consuming data processing. In this paper, we propose video similarity measurement for sports video categorization using expected value to average distance of video frames without the threshold. The distance between each sequence of the training and test videos was determined by comparing each sampling frame of the training videos with all sampling frames of the test videos and averaged by the expected value. In addition, each frame was represented with the color histogram to help enhance feature reduction, resulting in faster data processing. After that, the nearest neighbor (NN) classifier was applied to compare the similarity of the videos. Our experimental results show that this approach can achieve 97.0% accuracy in sports video similarity measurement.

Higher-order random projection for tensor object recognition

In this paper, the higher-order random projection (HORP) is proposed to directly project the higher-order tensor object from high-dimensional space to low-dimensional space for recognition task. In traditional random projection framework, the projection matrix does not depend on the training data hence it can avoid the principal classification problems such as over-fitting, Small Sample Size (SSS), and singularity problems. However, the tensor object must be transformed to vectors before projection. In this way, the size of projection matrix will depend on the product of all dimensions in all orders, that is very large and consumes lots of memory and computation time to process. Instead of the traditional projection, our method uses n-mode projection for a tensor object directly, which applies the random projection matrices to matrix unfolding in each mode simultaneously. Thus, the size of each projection matrix will depend on only a dimension of each order. The memory and computation time of this method will be substantially reduced. After projection, we investigate the results of HORP by the nearest neighbor classifier. Our experiments on well-known face databases demonstrate the significant of our proposed method.

2DPCA Feature Selection Using Mutual Information

In two-dimensional principal component analysis (2DPCA), 2D face image matrices do not need to be previously transformed into a vector. In this way, the image covariance matrix can be better estimated, compared to the old fashion. The feature is derived from eigenvectors corresponding to the largest eigenvalues of the image covariance matrix for data of all classes. Normally, the number of the largest eigenvalues is selected manually for obtaining the optimal feature matrix. In this paper, we propose a novel method for feature selection in 2DPCA, based on mutual information concept for automatically selecting the number of the largest eigenvalues. The non-parametric quadratic mutual information between class labels and features is used as a selection criterion. This does not only allows reducing of the dimension of feature matrix but also obtaining the good recognition accuracy. Experimental results on Yale face database showed an efficient of our proposed method.

Face Hallucination Using Bilateral-Projection-Based Two-Dimensional Principal Component Analysis

In this paper, we propose a new super-resolution face hallucination method based on Bilateral-projection-based Two-Dimensional Principal Component Analysis (B2DPCA). Firstly, the high-resolution (HR) face image and its corresponding low-resolution (LR) face image are projected to the HR and LR B2DPCA feature spaces, respectively. In these spaces, the linear mixing relationship between HR and LR feature is estimated from a training set. For reconstructing the HR image from the observed LR image, the LR image is firstly projected to LR feature space and then mapped to HR feature. Finally, the HR feature is reconstructed to the HR face image. Experiments on the well-known face databases show that the performance of our proposed method. The resolution and quality of the hallucinated face images are greatly enhanced over the LR ones, which is very helpful for human recognition.