Preface

An invariant object recognition system needs to be able to recognise the object under any usual *a priori* defined distortions such as translation, scaling and in‐plane and out‐ of‐plane rotation. Ideally, the system should be able to recognise (detect and classify) any complex scene of objects even within background clutter noise. This problem is a very complex and difficult one. In this book, we present recent advances towards achieving fully‐robust object recognition. In the book's chapters' cutting edge recent research is discussed in details for all the readers with a core or wider interest in this area.

In section 1, the relation and importance of object recognition in the cognitive processes of humans and animals is described as well as how human‐ and animal‐like cognitive processes can be used for the design of biologically‐inspired object recognition systems. Chapter 1 discusses about the neurophysiopathology of attention, learning and memory. Then, it discusses about object recognition, and a novel object recognition test and its role in building an experimental model of Alzheimer's disease. Chapter 2 discusses about episodic human memory and episodic‐like animal memory. Then, it discusses about object recognition and a novel object recognition task which can be used as an experimental tool for investigating full episodic memory in different animal species. Chapter 3 presents the performance analysis of the biologically‐ inspired modified‐hybrid optical neural network object recognition system within cluttered scenes. The system's biologically‐inspired hybrid design is analysed and is shown to combine a knowledge representation unit being the optical correlator block with a knowledge learning unit being the NNET block. Several experiments were conducted for testing the system's problem solving abilities as well as its performance in recognising multiple objects of the same or different classes within cluttered scenes.

In section 2, we discuss about colour processing and how it can be used to improve object recognition. Chapter 4 reviews the current state‐of‐the‐art research about the specific role of colour information in object recognition. Then, it investigates the role of colour in the recognition of colour and non‐colour diagnostic objects at different levels of the brain's visual processing.

In literature, we can identify two main categories of object recognition systems. The first category consists of linear combinatorial type filters. The second category consists of pure neural modelling approaches. In section 3, we discuss about those two categories of optical correlators and of artificial neural networks, respectively. Chapter 5 presents an iterative approach for synthesizing adaptive composite correlation filters for object recognition. The approach can be used for improving the quality of a simple composite filter in terms of quality metrics using all available information about the true‐class object to be recognised and false‐class objects to be rejected such as the background. Two different filters employing this iterative approach are described. First, an adaptive constrained filter is described which optimises its class discrimination properties, and, second, an adaptive unconstrained composite filter is described which optimises its properties with respect to the average correlation height (ACH), average correlation energy (ACE) and average similarity measure (ASM). Chapter 6 presents a method of integrating image features from the object's contour, its type of curvature or topographical surface information and depth information from a stereo camera, and then after being concatenated form an invariant vector descriptor which is input to a Fuzzy ARTMAP artificial neural network for learning and recognition purposes. Experimental results are discussed when using a single contour vector description (BOF), a combination of surface information vector (SFS) with BOF, and the full concatenated vector of BOF+SFS+Depth.

Preface XI

UK

UK

UK

**Ioannis Kypraios**

Dept. of Engineering and IT, at ICTM, London,

The Editor would like to acknowledge everyone that supported and actively or with their Prayers helped the successful completion of the book's publishing process stages. Special thanks to Ms Sasa Leporic InTech 's Publishing Process Manager for her

APEM Computing Labs (Remote Sensing –R&D Division), Centre for Innovation & Enterprise, Oxford University, Begbroke Science Park, Begbroke Hill, Oxfordshire,

School of Engineering and Design, University of Sussex, Falmer, Brighton,

unceasingly support and great help throughout the publication process.

In section 4, we present two different applications of object recognition with still images and with video sequences. Chapter 7 presents an application of object recognition for the discrimination of modern coins into several hundreds of different classes, and the identification of hand‐made ancient coins. Modern coins are acquired by a machine vision system for coin sorting but for ancient coins a scanner and camera devices are considered. In particularly, the use of a 3D acquisition device and 3D models of ancient coins are discussed. Different methods of segmentation are discussed for modern and ancient coins. Two main methods for classification are compared, one based on matching edge features in log‐polar space and a second method based on an eigenspace representation. For the identification of coins features extracted from the edge of a coin and from the Fourier domain representation of the coin contour are used, and a Bayesian fusion of coin sides is studied. Improvement by 3D analysis and modelling is also presented. Results are discussed for all considered datasets and methods. Chapter 8 presents an application of non‐rigid objects recognition in video sequences. An approach for recognising human action using spatiotemporal interest points (STIPs) is described. The STIPs are detected by employing different detectors. Several motion analysis techniques are presented, such as activity function, human body interest regions, and spatiotemporal boxes. Those techniques can be applied on a set of detected STIPs as an effective way of action representation. Several motion classification algorithms are discussed, such as support vector machines (SVM), probabilistic latent semantic analysis (pLSA) and others, and a proposed by the authors algorithm based on unsupervised k‐means clustering algorithm. The proposed algorithm is compared with existing algorithms by being tested with the KTH human action database.

The Editor would like to acknowledge everyone that supported and actively or with their Prayers helped the successful completion of the book's publishing process stages. Special thanks to Ms Sasa Leporic InTech 's Publishing Process Manager for her unceasingly support and great help throughout the publication process.

X Preface

of pure neural modelling approaches. In section 3, we discuss about those two categories of optical correlators and of artificial neural networks, respectively. Chapter 5 presents an iterative approach for synthesizing adaptive composite correlation filters for object recognition. The approach can be used for improving the quality of a simple composite filter in terms of quality metrics using all available information about the true‐class object to be recognised and false‐class objects to be rejected such as the background. Two different filters employing this iterative approach are described. First, an adaptive constrained filter is described which optimises its class discrimination properties, and, second, an adaptive unconstrained composite filter is described which optimises its properties with respect to the average correlation height (ACH), average correlation energy (ACE) and average similarity measure (ASM). Chapter 6 presents a method of integrating image features from the object's contour, its type of curvature or topographical surface information and depth information from a stereo camera, and then after being concatenated form an invariant vector descriptor which is input to a Fuzzy ARTMAP artificial neural network for learning and recognition purposes. Experimental results are discussed when using a single contour vector description (BOF), a combination of surface information vector (SFS) with BOF,

In section 4, we present two different applications of object recognition with still images and with video sequences. Chapter 7 presents an application of object recognition for the discrimination of modern coins into several hundreds of different classes, and the identification of hand‐made ancient coins. Modern coins are acquired by a machine vision system for coin sorting but for ancient coins a scanner and camera devices are considered. In particularly, the use of a 3D acquisition device and 3D models of ancient coins are discussed. Different methods of segmentation are discussed for modern and ancient coins. Two main methods for classification are compared, one based on matching edge features in log‐polar space and a second method based on an eigenspace representation. For the identification of coins features extracted from the edge of a coin and from the Fourier domain representation of the coin contour are used, and a Bayesian fusion of coin sides is studied. Improvement by 3D analysis and modelling is also presented. Results are discussed for all considered datasets and methods. Chapter 8 presents an application of non‐rigid objects recognition in video sequences. An approach for recognising human action using spatiotemporal interest points (STIPs) is described. The STIPs are detected by employing different detectors. Several motion analysis techniques are presented, such as activity function, human body interest regions, and spatiotemporal boxes. Those techniques can be applied on a set of detected STIPs as an effective way of action representation. Several motion classification algorithms are discussed, such as support vector machines (SVM), probabilistic latent semantic analysis (pLSA) and others, and a proposed by the authors algorithm based on unsupervised k‐means clustering algorithm. The proposed algorithm is compared with existing algorithms by being

and the full concatenated vector of BOF+SFS+Depth.

tested with the KTH human action database.
