Catálogo de publicaciones - libros

In the paper the multilevel probabilistic approach to handprinted form recognition is described. The form recognition is decomposed into three levels: character recognition, word recognition and form contents recognition. On the word and form contents level the probabilistic lexicons are available. The decision on the word level is performed using probabilistic properties of character classifier and the contents of probabilistic lexicon. The novel approach to combining these two sources of information about classes (words) probabilities is proposed, which is based on lexicons and accuracy assessment of local character classifiers. Some experimental results and examples of practical applications of recognition method are also briefly described.

In the paper the method of handprinted word recognition is described, which combines statistical lexical language model and character classifier properties in order to improve the recognition accuracy. The statistical lexical model determines the conditional probabilities of letters succession in the language. For some letters in polish language only very small subset of successors appears with significant conditional probability. If the confidence of predecessor recognition is assessed as high then the recognition of successor can be reliably supported by utilizing probabilistic lexical properties. In contrast to many other approaches, the method is not based on lexicons, so it can be used in these cases where the exhaustive lexicon is not available or its usage is inefficient, e.g. due to great number of elements.

The paper presents the speaker recognition problem in the background of voice transmission via Internet. The Gaussian Mixture Models (GMM) classification, the voice feature extraction, the Internet speech transmission standards and the packet loss simulation methodology applied in the tested system were overviewed. Speaker identification scores obtained for the tested GMM based text-dependent system has revealed a minor significance of packet loss problem in this aspect.

The important element connected with today’s speech recognition/ synthesis systems is the speech database — the set of fully annotated wavefiles. Since the manual segmentation of speech is a very time-consuming task, the automatic segmentation algorithms are needed. However, the manual segmentation still outperforms the automatic one and at the same time the quality of resulting synthetic voice highly depends on the accuracy of the phonetic segmentation. In this paper we concentrate on a semi-automatic approach, in which a human expert, unlike in the common approach, manually allocates the selected boundaries prior to the automatic segmentation of the rest of the corpus. In the paper we quest for the appropriate strategy for an expert. We check if locating some boundary classes influence the rest of the annotations. It is done for two difierent quality measures.

This paper describes the original method of user’s identification. The method bases on the fusion of Wavelet Transform (WT) and Hidden Markov Models (HMM), which is used for three parts of the face (eyes, nose, and mouth) separately.

One of the major difficulties encountered in the development of face image processing algorithms, is the possible presence of occlusions that hide part of the face images to be processed. Typical examples of facial occlusions include sunglasses, beards, hats and scarves. In our work we address the problem of restoring the overall shape of faces given only the shape presentation of a small part of the face. For this purpose a novel technique which utilizes combination of Hopfield and Multi-Layer Perceptron (MLP) neural networks was used. According to the experimental results it is possible to recover with reasonable accuracy the overall shape of faces even in the case where a substantial part of the shape of a given face is not visible. The presented technique could form the basis for developing face image processing systems capable of dealing with occluded faces.

This paper presents comparison of three minutiae matching techniques, i.e. Hough transform, global star method and orientation correlation. Short description of the pre-processing stage based on filtering, thinning and minutiae extraction is presented. The investigations are performed with a high quality fingerprint scanner.

In the article, the system for fitting the face-shape model to the image is described. Two model fitting approaches have been used: ACM and ASM. ACM has been applied in order to create a face-contours base of models. Starting with that base, an ASM model has been computed, which then has been used to obtain the face-shape descriptions from the provided images. ASM gives the explicit and structured shape description from the source image. Implementation of the ASM-based recognition system has been described and its performance evaluated. Some conclusions related to the choice of the local support are given, based on the experimental validation of the ASM method.

In this paper, a face recognition algorithm based on statistical model of Active Shape (ASM) is presented. A 31 degree-of-freedom shape model was used. The model was derived from a set of 183 faces shapes and named the learning set. Criteria of selection of face to model classifiers were evaluated. Classification was implemented in the shape space, in its Principal Component Analysis (PCA) and Multiple Discriminant Analysis (MDA) transformations. In the shape space the metrics were used. metric was used in PCA and MDA spaces as well. The results were based on experiments carried out on the set of 651 images of eight persons. Further proceedings in the case of ambiguous classification results were suggested.

In the article new method of face recognition is considered. It exploits two well-known approaches namely DCT and LDA. Using LDA on selected spectral components of the DCT better separation of classes can be achieved. Scaling problem of the face images was addressed and appropriate solution proposed. Experiments on the ORL [] database of faces were carried out. Results were compared with the individual DCT approach and one of the most frequently used approach nowadays: PCA+LDA.