Volume 26, Nº 2 (2016)

Linear transformation of data in multidimensional feature space based on Fisher’s criterion is considered. The case of two classes with arbitrary distributions is studied. We derived expressions for recurrent calculation of weight vectors which form new features. Example offered shows that the newly found features which represent the data more accurately make it possible to achieve linear separability of classes which remains impossible using the technique of principal components and the classic Fisher’s linear discriminant.

Method of fuzzy boosting providing iterative weak classifiers selection and their quasi-linear composition construction is presented. The method is based on the combination of boosting and fuzzy integrating techniques, when at each step of boosting weak classifiers are combined by Choquet fuzzy integral. In the proposed FuzzyBoost algorithm 2-additive fuzzy measures were used, and method for their estimation was proposed. Although detailed theoretical verification of proposed algorithm is still absent, the experimental results, made on simulated data models, demonstrate that in the case of complex decision boundaries FuzzyBoost significantly outperforms AdaBoost.

There are many modern devices are used to create digital images. These devices use optical effects to create images. Therefore, the image quality depends on quality of optical sensors. Because of the limits of technology, these sensors cannot reconstruct the images perfectly, and always include some defects. One from these defects is noise. The noise reduces image quality and result of image processing. The image noises can be classified into some types: Gaussian noise, Poisson noise, speckle noise and so on. Depending on particular noises, we have efficient methods to remove them. There is no existing a universal method to remove all noises effectively. In this paper, we proposed a method to remove a noise that is popular in biomedicine. This noise can be considered as a combination of Gaussian and Poisson noises. Our method is based on the total variation of an image intensity (brightness) function.

A progressive framework is proposed for dense stereo matching to solve problems caused by weaktexture and occlusion in this paper. The main idea is that disparity is extracted progressively, from coarse to fine, from sparse to dense. First, a coarse disparity map is obtained by the segment-based pre-matching method, in which horizontal and vertical segment matching are performed in parallel and pre-matching results are merged to preserve more details. Second, disparity diffusion is performed to roughly estimate disparity values for miss-matched points. Third, a probabilistic approach is used for disparity refinement, taking into account stereo prior, image likehood and disparity smoothness. Experiments are made on the Middlebury benchmark to demostrate the effectiveness of the proposed algorithm.

In this paper we present how adaptable learned models of graphical models are and how they can be used for classification tasks of 3D laser point clouds with different distributions and density. In order to model the contextual information we use a pair-wise conditional random field and an adaptive graph down-sampling method based on voxel grids. As feature we apply the rotation invariant histogram-of-oriented-residuals operator to describe the local point cloud distribution. We validate the approach with data collected from different laser range finders with varying point cloud distribution and density. Our experiments imply, that conditional random field models learned from one dataset can be applied to another dataset without a significant loss of precision.

Image segmentation is one of many fundamental problems in computer vision. The need to divide an image to a number of classes is often a part of a system that uses image processing methods. Therefore, lots of methods were developed that are based on different approaches. The image segmentation could be classified with respect to many criteria. One such a criterion is based on the degree of allowed interactivity. The interactivity could be of several types—interactive initialization, interaction while the computation is running or manual refinement of achieved results, for example. Especially the precise initialization plays an important role in many methods. Therefore the possibility to initialize the method manually is often invaluable advantage and information obtained this way could be the difference between good and poor results. Unfortunately, in many cases it is not possible to initialize a method manually and the process needs to be automated. In this paper, an approach for such an automation is presented. It is based on shortest paths in a graph and deriving an area of influence for each obtained seed point. This method is called shortest path basins.

We address the problem of recognizing the object with distinctive edge features. For this purpose, a recognition approach based on local edge features is presented. First the edge features are detected in each image, and then its descriptor is computed to find the match features. Each match will give a vote with location, scale and orientation of the object. The recognition result can be found in the densest position in the vote space. Experimental results show that the presented method is robust and effective to the object with distinctive edge features.

Presently there already existed a few human ear image databases, which have been very influential in advancing the research on ear recognition. However no ear video databases are available. In this paper, we introduce the construction and basic content of the Chinese Ear Video Database (CEVD) and some primary evaluation results on it. The CEVD consists of 3600 ear video segments collected from 120 subjects. All the video segments in the database were collected in specially designed environment with three principal variations of illumination condition, viewing angle and interference. Compared with other public ear database, CEVD can not only be used for image-based applications but also video-based applications. In this paper we introduce the database and describe the collecting procedure. It excels in its large-scale and variation modes and is expected to have positive impact on the development and evaluation of ear recognition algorithms. This paper also gives experiment results of improved CamShift and AdaBoost algorithm on the database.

In this paper, a fast and reliable method for hand detection based on continuous skeletons approach is presented. It demonstrates real-time working speed and high detection accuracy (3–5% both FAR and FRR) on a large dataset (50 persons, 80 videos, 2322 frames). These make it suitable for use as a part of modern hand identification systems including mobile ones. Overall, the study shows that continuous skeletons approach can be used as prior for object and background color models in segmentation methods with supervised learning (e.g., interactive segmentation with seeds or abounding box).

The authors propose a new face recognition system with an evaluation function using feature points. The feature points are detected automatically by Milborrow’s Stasm software. Before recognition, rotation compensation and size normalization are applied to the feature points. The main method is to calculate the squared error between the registered face and the input face as to length of a characteristic pair of feature points on face. The False Rejection Rate (FRR) for the registered and input face of the same person, and the False Acceptance Rate (FAR) for the registered face and a different person’s input face are evaluated. The input is a video sequence. Stable recognition is obtained with small FRR and FAR for the video of a period of 0.5 s.

A method for determining the pupil boundary in the image of eye is proposed. The method is based on image binarization followed by a search of the pupil as one of the connectivity components. The pupil boundary is determined as a part of boundary of the connectivity component. Hough transform is used for separating pupil in the case of its merging in one connectivity component with other objects, as well as to verify the likelihood of solution.

This paper addresses the problem of association rule extraction. To extract quantitative association rules from given sets of observations, a stochastic method is proposed. The developed method improves the reliability and interpretability of recognition models based on association rules, employs the stochastic approach to search through various combinations of sets of elements in association rules, and uses a priori information about the informativity of intervals of feature values. A system of criteria for estimating association rules is developed that can be used to automate the analysis of properties and to compare various models based on association rules when solving pattern recognition problems.

Needs and new technologies always inspire people to make new ways to interact with machines. This interaction can be for a specific purpose or a framework which can be applied to many applications. Sign language recognition is a very important area where an easiness in interaction with human or machine will help a lot of people. At this time, India has 2.8M people who can’t speak or can’t hear properly. This paper targets Indian sign recognition area based on dynamic hand gesture recognition techniques in real-time scenario. The captured video was converted to HSV color space for pre-processing and then segmentation was done based on skin pixels. Also Depth information was used in parallel to get more accurate results. Hu-Moments and motion trajectory were extracted from the image frames and the classification of gestures was done by Support Vector Machine. The system was tested with webcam as well as with MS Kinect. This type of system would be helpful in teaching and communication of hearing impaired persons.

The problem of speaker recognition from a given set of speakers for any language and any context is considered. A database of Russian numerals that contains speech segments from 216 men and 177 women, each of whom spoke from 400 to 800 words, is used for recognition. Speech has been recorded on different types of microphones in different rooms at the natural noise level. Recognition is based on solutions of the inverse problem of finding the voice excitation pulse shape for each pitch period by the known speech segment. The pulse shape is defined as the inverse Fourier transform of the regularized ratio of speech signal spectra at the intervals of the open and closed glottis. Recognition is carried out by ten parameters: the pitch period, the open glottis interval duration, times when the source amplitude is maximum, minimum, or zero, the amplitude ratio for the minimum and maximum source pulses, three decomposition ratios of the source function by the principal component method, and the vowel duration. In such a recognition procedure, in the case of the utterance of a word that contains one vowel, the false reject rate (FRR) for men is 1.7–5.4%, and the false acceptance rate (FAR) is 5.4–7.1%. For women FRR = 2–5.2% and FAR = 5.2–6.3%. The recognition error decreases with an increasing number of vowels in the speech signal. At 10 vowels, for men FRR = 0.05–0.2% and FAR = 0.07–0.8%, and for women FRR = 0.09–0.2% and FAR = 0.17–2.1%.