Pavel Král

Automatic face recognition system based on the SIFT features

We proposed and implemented a new face corpus creation algorithm.We created a new facial corpus from the data of the Czech News Agency.We evaluated a novel face recognition method, the SIFT based Kepenekci approach.We proposed and evaluated two novel confidence measure techniques.We proposed, implemented and evaluated the fully automatic face recognition system. The main goal of this paper is to propose and implement an experimental fully automatic face recognition system which will be used to annotate photographs during insertion into a database. Its main strength is to successfully process photos of a great number of different individuals taken in a totally uncontrolled environment. The system is available for research purposes for free. It uses our previously proposed SIFT based Kepenekci approach for the face recognition, because it outperforms a number of efficient face recognition approaches on three large standard corpora (namely FERET, AR and LFW). The next goal is proposing a new corpus creation algorithm that extracts the faces from the database and creates a facial corpus. We show that this algorithm is beneficial in a preprocessing step of our system in order to create good quality face models. We further compare the performance of our SIFT based Kepenekci approach with the original Kepenekci method on the created corpus. This comparison proves that our approach significantly outperforms the original one. The last goal is to propose two novel supervised confidence measure methods based on a posterior class probability and a multi-layer perceptron to identify incorrectly recognized faces. These faces are then removed from the recognition results. We experimentally validated that the proposed confidence measures are very efficient and thus suitable for our task.

Evaluation of the Document Classification Approaches

This paper deals with one class automatic document classification. Five feature selection methods and three classifiers are evaluated on a Czech corpus in order to build an efficient Czech document classification system. Lemmatization and POS tagging are used for a precise representation of the Czech documents. We demonstrated, that POS tag filtering is very important, while the lemmatization plays a marginal role for classification.We also showed that Maximum Entropy and Support Vector Machines are very robust to the feature vector size and outperform significantly the Naive Bayes classifier from the view point of the classification accuracy. The best classification accuracy is about 90% which is enough for an application for the Czech News Agency, our commercial partner.

Automatically Detected Feature Positions for LBP Based Face Recognition

This paper presents a novel approach for automatic face recognition based on the Local Binary Patterns (LBP). One drawback of the current LBP based methods is that the feature positions are fixed and thus do not reflect the properties of the particular images. We propose to solve this issue by a method that automatically detects feature positions in the image. These key-points are determined using the Gabor wavelet transform and k-means clustering algorithm. The proposed method is evaluated on two corpora: AT&T Database of Faces and our Czech News Agency (CTK) dataset containing uncontrolled face images. The recognition rate on the first dataset is 99.5% which represents 2.5% improvement compared to the original LBP method. The best recognition rate obtained on the CTK corpus is 59.1% whereas the original LBP method reaches only 38.1%.

Statistical Language and Speech Processing

With the development of the web of data, recent statistical, data-to-text generation approaches have focused on mapping data (e.g., database records or knowledge-base (KB) triples) to natural language. In contrast to previous grammar-based approaches, this more recent work systematically eschews syntax and learns a direct mapping between meaning representations and natural language. By contrast, I argue that an explicit model of syntax can help support NLG in several ways. Based on case studies drawn from KB-to-text generation, I show that syntax can be used to support supervised training with little training data; to ensure domain portability; and to improve statistical hypertagging.

Novel Matching Methods for Automatic Face Recognition Using SIFT

The object of interest of this paper is Automatic Face Recognition (AFR). The usual methods need a labeled corpus and the number of training examples plays a crucial role for the recognition accuracy. Unfortunately, the corpus creation is very expensive and time consuming task. Therefore, the motivation of this work is to propose and implement new AFR approaches that could solve this issue and perform well also with few training examples. Our approaches extend the successful method based on the Scale Invariant Feature Transform (SIFT) proposed by Aly. We propose and evaluate two methods: the Lenc-Kral matching and the SIFT based Kepenekci approach [7]. Our approaches are evaluated on two face data-sets: the ORL database and the Czech News Agency (CTK) corpus. We experimentally show that the proposed approaches significantly outperform the baseline Aly method on both corpora.

Automatic Dialog Acts Recognition Based on Sentence Structure

This paper deals with automatic dialog acts (DAs) recognition in Czech. Our work focuses on two applications: a multimodal reservation system and an animated talking head for hearing-impaired people. In that context, we consider the following DAs: statements, orders, investigation questions and other questions. The main goal of this paper is to propose, implement and evaluate new approaches to automatic DAs recognition based on sentence structure and prosody. Our system is tested on a Czech corpus that simulates a task of train tickets reservation. With lexical-only information, the classification accuracy is 91%. We proposed two methods to include sentence structure information, which respectively give 94% and 95%. When prosodic information is further considered, the recognition accuracy reaches 96%

Deep Neural Networks for Czech Multi-label Document Classification

This paper is focused on automatic multi-label document classification of Czech text documents. The current approaches usually use some pre-processing which can have negative impact (loss of information, additional implementation work, etc). Therefore, we would like to omit it and use deep neural networks that learn from simple features. This choice was motivated by their successful usage in many other machine learning fields. Two different networks are compared: the first one is a standard multi-layer perceptron, while the second one is a popular convolutional network. The experiments on a Czech newspaper corpus show that both networks significantly outperform baseline method which uses a rich set of features with maximum entropy classifier. We have also shown that convolutional network gives the best results.

Neural network acoustic model with decision tree clustered triphones

This article tries to compare the performance of neural network and Gaussian mixture acoustic models (GMMs). We argue that using a multi layer perceptron as an emission probability estimator in hidden Markov model based automatic speech recognition can lead to better results than when the more traditional Gaussian mixtures are applied. We present a solution on how to model triphone phonetic units with neural networks and we show that this also leads to better performance in comparison with GMMs. The superior performance of the neural networks comes at a cost of extremely long training times.

Multi-label Document Classification in Czech

This paper deals with multi-label automatic document classification in the context of a real application for the Czech news agency. The main goal of this work is to compare and evaluate three most promising multi-label document classification approaches on a Czech language. We show that the simple method based on a meta-classifier proposes by Zhu at al. outperforms significantly the other approaches. The classification error rate improvement is about 13%. The Czech document corpus is available for research purposes for free which is another contribution of this work.

Local binary pattern based face recognition with automatically detected fiducial points

This paper deals with automatic face recognition in the context of a real application for person identification developed for the Czech News Agency (y CTK) . We focus on popular Local Binary Patterns (LPBs) that are frequently used in this field with high recognition accuracy. One drawback of current LBP based methods is that the positions and number of the fiducial points are fixed. These points thus do not reflect the properties of a particular image whereas we believe it is beneficial to identify the most representative ones. The main contribution consists in proposing and comparing several LBP-based approaches that detect such points fully automatically. We use a set of Gabor filters for this task. Local extrema in the filter responses are detected and then used as the feature points. The number of points is further reduced by a clustering algorithm. Our approaches also differ from the other ones in the matching procedure. The proposed methods are evaluated on three standard corpora: ORL, FERET, AR face database and our y CTK dataset containing uncontrolled face images. Recognition results clearly show high quality of the proposed approaches that outperform significantly the baseline LBP approach on all corpora. The benefits of our methods are particularly evident in the case of real non-controlled images (y CTK corpus) where the accuracy is increased by more than 20% in absolute value.