Human and Machine Perception

PREFACE

SPONSORING INSTITUTIONS

CONTENTS

In this chapter we discuss some issues on mobile robot perception from the point of view of the structure of vision systems. Through some experiences and results from research projects we will approach the problem from the point of view of purposive vision, with particular regard to less conventional sensors and systems, such as omnidirectional and hybrid systems.

The chapter introduces new entropy measures that use the image information content such as grey levels and their topological distribution in the image domain in order to perform the classification of the image itself. The main aim of the chapter is to study the role of the image entropy in perceptual tasks and to compare the proposed approach with others well-known methods. Experiments have been carried out on medical images (mammograms) due to their variability and complexity. The image entropy approach seems to work quite well and it is less time-consuming if compared with the other methods.

Biomedical signals carry important information about the behavior of the living systems under studying. A proper processing of these signals allows in many instances to obtain useful physiological and clinical information. Many advanced algorithms of signal and image processing have recently been introduced in such an advanced area of research and therefore important selective information is obtainable even in presence of strong sources of noise or low signal/noise ratio. Traditional stationary signal analysis together with innovative methods of investigation of dynamical properties of biological systems and signals in second-order or in higher-order approaches (i.e., in time-frequency, time-variant and time-scale analysis, as well as in non linear dynamics analysis) provide a wide variety of even complex processing tools for information enhancement procedures. Another important innovative aspect is also remarked: the integration between signal processing and modeling of the relevant biological systems is capable to directly attribute patho-physiological meaning to the parameters obtained from the processing and viceversa the modeling fitting could certainly be improved by taking into account the results from signal processing procedure. Such an integration process could comprehend parameters and observations detected at different scales, at different organs and with different modalities. This approach is reputed promising for obtaining an olistic view of the patient rather than an atomistic one which considers the whole as a simple sum of the single component parts.

The binding problem stems from our difficulty in understanding how the brain is able to integrate different features, associated to the outputs of different detectors, into the unitary perception of a whole object. We analyze critically the claim that binding is produced by a synchronization of neural oscillations. To this regard we evidence how this proposal is untenable in its original form. We underline, however, that a suitable modification of it, taking into account the quantum nature of physical processes occurring within the brain, can open the way toward a more realistic solution of binding problem and a better explanation of most phenomena evidenced by psychology of visual perception.

Every time a user links up to a web site, the server keeps track of all the actions accomplished in a log file. What is captured is the “click flow” (clickstream) of the mouse and the keys used by the user during the navigation inside the site. Usually every click of the mouse corresponds to the viewing of a web page. Therefore, we can define the click-stream as the sequence of the web pages requested. The objective of this chapter is to show how web clickstream data can be used to understand the most likely paths of navigation in a web site, with the aim of predicting, possibly on-line, which pages will be seen, having seen a specific path of other pages before. Such analysis can be very useful to understand, for instance, what is the probability of seeing a page of interest (such as the buying page in an e-commerce site) coming from another page. Or what is the probability of entering (or exiting) the web site from any particular page. From a methodological viewpoint, our aim is to present new associative models, obtained by means of statistical graphical Markov models, and compare them with classical association rules, direct or embodied in classification tree models. More specifically, as web pages are ordered in time, we shall consider sequence rules.

This chapter presents a face recognition method comparing a 3D facial model, to a gallery of previously enrolled faces. The geometrical features in the input mesh are represented by a bidimensional matrix, the normal map, storing local curvature data as the rgb components of a color image. The robustness of the proposed approach to facial expressions has been improved by a weighted mask automatically calculated for every subject in the gallery using a set of specific expressive variations. As comparison between normal maps is executed in a 2D space, the computational cost of this method is low. We present the results of our method on a 3d database of human faces, featuring different races, sex, ages, and expressions.

The chapter introduces new entropy measures that use the image information content such as grey levels and their topological distribution in the image domain in order to perform the classification of the image itself. The main aim of the chapter is to study the role of the image entropy in perceptual tasks and to compare the proposed approach with others well-known methods. Experiments have been carried out on medical images (mammograms) due to their variability and complexity. The image entropy approach seems to work quite well and it is less time-consuming if compared with the other methods.

This chapter deals with the classification of complex objects by morphological shape analysis and elliptical Fourier descriptors. An unsupervised method has been proposed to identify components with specific shapes by a simple edge detector and to classify them via the description of their contours. A particular application has been arranged in order to evaluate the goodness of this approach when discriminating between normal and pathological human megakaryocytes. Alterations in these cells can occur in many pathological processes and in such cases the pattern, size and shape of the cytoplasm and/or of the nucleus are extremely varied.

Due to the overwhelming amount of information contained in the images produced by an information visualization tool, an important requirement for any visualization technique is that they should support rapid, accurate, and effortless visual exploration. We address this goal converting information into a graphical representation that takes better advantage of the human visual attention mechanisms.

In this paper, we provide some examples of integration of human and machine perception, derived from a specific research project that aims at emulating the human vision system.

This chapter is about vision to make systems more autonomous. We parallel two aspects of the current evolution of System-Perception: more ambitious yet coherent tasks tightly rely on more abstract description and control. As vision is likely to be a major and complex sensory modality for machines as it is for most animals, we concentrate our development on it. In the first part we show how thinking to systems helped to better pose vision problems and solve them in a useful manner. That is the “active vision” trend that we explain and illustrate. Along the same line, the necessity for anticipation further shows, leading to a first definition of “situated vision”. The second part deals with how to design systems able to achieve such vision. We show from a few examples how architectural descriptions evolve and better fit important features to grasp –a model– in view of more efficient control towards intelligence. Inner communication flows are better be controlled than local tasks that should be assumed completed efficiently enough in all cases. We conclude with a plausible sketch of a system to be experimented on in situations that require some autonomy.

A lot of importance has been given to the security problems in the last years. Particularly, secure access in reserved area with consentient people. Face Recognition works well providing high recognition rate with consentient/non-consentient people. However we have to deal with occlusions, such as sunglasses or scarfs. In this chapter a new face recognition method, namely ROF, is proposed, it is based on IFS (Iterated Function Systems) theory, in which the information used for the indexing and recognition task is made local, in order to render the method robust with respect to possible occlusions. The distribution of similarities in the face image is exploited as a signature for the identity of the subject. Besides it has been investigated on the amount of informations provided by each component of the face, first independently and then all together.