Michał Dolecki

An application of chain code-based local descriptor and its extension to face recognition

Local descriptors are widely used technique of feature extraction to obtain information about both local and global properties of an object. Here, we discuss an application of the Chain Code-Based Local Descriptor to face recognition by focusing on various datasets and considering different variants of this description method. We augment the generic form of the descriptor by adding a possibility of grouping pixels into blocks, i.e., effectively describing larger neighborhoods. The results of experiments show the efficiency of the approach. We demonstrate that the obtained results are comparable or even better than those delivered by other important algorithms in the class of methods based on the Bag-of-Visual-Words paradigm. An extension of Chain Code-Based Local Descriptor (CCBLD) is proposed.CCBLD is applied to face recognition task.Bag-of-Visual-Words paradigm is realized through the dictionary of chain-codes.Test results show that CCBLD is comparable or outperforms other local descriptors.The approach is tested using CAS-PEAL, ColorFERET, FG-NET, and other datasets.

Utility functions as aggregation functions in face recognition

Face recognition by computers in recent years has been a topic of intensive studies. In this problem, we witness several challenges: one has to cope with large data sets, solve problems of data extraction, and deal with poor quality of images caused by e.g., poor lighting of the subject. There have been a lot of algorithms and classifiers developed, which are aimed at recognizing faces of individuals. In this paper, we present a novel classification method, which involves a collection of classifiers with a certain utility function regarded as an aggregation operator. The nearest neighbor method with various similarity measures is used as a generic classifier for selected face areas. The main task is to assign photos of a person to one of the classes of image present in the available database. This problem is similar to the decision-making process with some evident analogies. If in face recognition, a single classifier is being used, the problem becomes similar to the one of decision-making with a single criterion. When having several classifiers, the problem resembles a problem of a multi-criteria decision making. The second scenario requires an aggregation of the results produced by different classifiers. The paper presents the use of the utility function which is well-known in the decision-making theory as an aggregation operator applied to the results of various classifiers. The study is focused on the two-factor utility function and its variants.

Face recognition by humans performed on basis of linguistic descriptors and neural networks

In this study, we present a new approach to the problem of face classification, which relies on the linguistic description of the facial features. In this method, face descriptors are represented through the Analytic Hierarchy Process (AHP) and formalized as information granules. Moreover, neural networks are used to construct efficient classifiers. Furthermore, with usage of AHP we realize a transition from the linguistic description of the facial features to the vectors of numbers that are used by a neural network in the process of matching faces. The results of experiments demonstrate the potential applicability of our proposal to the forensic investigations. Finally, discussed are important aspects of constructing neural networks regarded as a vehicle to perform classification process.

Linguistic Descriptors and Analytic Hierarchy Process in Face Recognition Realized by Humans

In this paper, we discuss an application of the linguistic descriptions obtained directly from experts’ and treated as the votes when characterizing facial images to carry out face classification. Despite various automated face recognition techniques, the expert’s opinion plays a pivotal role in making classification decisions when recognizing faces, say in problems of suspect identification. Here, we analyze the impact of critical factors (e.g., a number of experts, voting schemes, distance functions) and their impact on the performance of classification schemes. The well-established Analytic Hierarchy Process (AHP) is used to quantify importance of linguistic descriptors in the process of face recognition by humans. As a result we produce realistic weights improving the accuracy of classification. Experimental results are presented including a number of parametric studies.

Linguistic descriptors and fuzzy sets in face recognition realized by humans

In this study, we present a new approach to the face retrieval and face classification problem, which exploits available experts knowledge and introduces a novel way of describing facial features. These features are described by manually assigned weights corresponding to membership grades with respect to the linguistic descriptors such as short, medium, or long. In the series of experiments, we also use weights produced by the Analytic Hierarchy Process aimed at producing saliences of facial cues. We identify a group of the most essential facial features.

Threshold Method of Detecting Long-Time TPM Synchronization

The phenomenon of neural networks synchronization by mutual learning can be used to construct key exchange protocol on an open channel. For security of this protocol it is important to minimize knowledge about synchronizing networks available to the potential attacker. The method presented herein permits evaluating the level of synchronization before it terminates. Subsequently, this research enables to assess the synchronizations, which are likely to be considered as long-time synchronizations. Once that occurs, it is preferable to launch another synchronization with the new selected weights as there is a high probability (as previously shown) that a new synchronization belongs to the short one.

DISTRIBUTION OF THE TREE PARITY MACHINE SYNCHRONIZATION TIME

Neural networks’ synchronization by mutual learning discovered and described by Kanter et al. [12] can be used to construct relatively secure cryptographic key exchange protocol in the open channel. This phenomenon based on simple mathematical operations, can be performed fast on a computer. The latter makes it competitive to the currently used cryptographic algorithms. An additional advantage is the easiness in system scaling by adjusting neutral network’s topology, what results in satisfactory level of security [24] despite different attack attempts [12, 15]. With the aid of previous experiments, it turns out that the above synchronization procedure is a stochastic process. Though the time needed to achieve compatible weights vectors in both partner networks depends on their topology, the histograms generated herein render similar distribution patterns. In this paper the simulations and the analysis of synchronizations’ time are performed to test whether these histograms comply with histograms of a particular well-known statistical distribution. As verified in this work, indeed they coincide with Poisson distribution. The corresponding parameters of the empirically established Poisson distribution are also estimated in this work. Evidently the calculation of such parameters permits to assess the probability of achieving both networks’ synchronization in a given time only upon resorting to the generated distribution tables. Thus, there is no necessity of redoing again time-consuming computer simulations.

Distance of the initial weights of tree parity machine drawn from different distributions

It is well-known that artificial neural networks have the ability to learn based on the provisions of new data. A special case of the so-called supervised learning is a mutual learning of two neural networks. This type of learning applied to a specific networks called Tree Parity Machines (abbreviated as TPM networks) leads to achieving consistent weight vectors in both of them. Such phenomenon is called a network synchronization and can be exploited while constructing cryptographic key exchange protocol. At the beginning of the learning process both networks have initialized weights values as random. The time needed to synchronize both networks depends on their initial weights values and the input vectors which are also randomly generated at each step of learning. In this paper the relationship between the distribution, from which the initial weights of the network are drawn, and their compatibility is discussed. In order to measure the initial compatibility of the weights, the modified Euclidean metric is invoked here. Such a tool permits to determine the compatibility of the network weights’ scaling in regard to the size of the network. The proper understanding of the latter permits in turn to compare TPM networks of various sizes. This paper contains the results of the simulation and their discussion in the context of the above mentioned issue.

The Impact of the TPM Weights Distribution on Network Synchronization Time

Two neural networks with randomly chosen initial weights may achieve the same weight vectors in the process of their mutual learning. This phenomenon is called a network synchronization, and can be used in cryptography to establish the keys for further communication. The time required to achieve consistent weights of networks depends on the initial similarity and on the size of the network. In the previous work related to this topic the weights in TPM networks are randomly chosen and no detailed research on used distribution is performed. This paper compares the synchronization time obtained for the network weights randomly chosen from either the uniform distribution or from the Gaussian distribution with different values of standard deviation. The synchronization time of the network is examined here as a function of different numbers of inputs and of various weights belonging to the intervals with varying sizes. The standard deviation of Gaussian distribution is selected depending on this interval size in order to compare networks with different weights intervals, which also constitutes a new approach for selecting the distribution’s parameters. The results of all analyzed networks are shown as a percentage of the synchronization time of a network with weights drawn from uniform distribution. The weights drawn from the Gaussian distribution with decreasing standard deviation have shorter synchronization time especially for a relatively small network.