A wide variety of approaches using neural networks for solving signal and image processing problems are presented. The principal focus is on multilayer perceptrons, which Masters calls multiple-layer feedforward networks (MLFN).
The book introduces problems of signal and image processing in the complex domain and proposes the use of complex-domain neural networks (networks of neurons that use complex numbers instead of real numbers). The algorithms are presented well and justified along with all the associated formulas. Source code of the algorithms and a complete working program using MLFN are included on a disk provided with the book.
The author reviews signal and image processing algorithms, including Gabor, Fourier, and Morlet wavelet transforms, convolution algorithms, and algorithms that operate in the frequency domain. The book is biased in favor of the use of complex numbers; the point is well made and supported with real-life examples.
The book is organized into ten chapters and an appendix:
The Role of Neural Networks in Signal and Image Processing
Neurons in the Complex Domain
Data Preparation for Neural Networks
Frequency-domain Techniques
Time/frequency Localization
Time/frequency Applications
Image Processing in the Frequency Domain
Moment-based Image Features
Tone/texture Descriptors
Using the MLFN Program
The appendix is “About the Program Disk.” In chapter 1, the author tries to examine some reasons for considering neural networks in the resolution of problems for which other solutions already exist, but he really presents some considerations about the selection and configuration of neural network models depending on the problem to be solved. Some examples of successful neural network solutions to real-world problems are listed, but the objective of justifying the use of neural networks to solve them is not achieved until the end of the book.
The main argument used in chapter 2 is the selection of an abstract representation of the problem, within a neural network, that reflects the nature of the data to be processed. Many physical phenomena, such as wavelets or phase transitions, are in the complex domain, and the methods commonly used to process signals bring the data into the realm of complex numbers. This chapter presents the construction of MLFNs based on neurons in the complex domain, that is, neurons with special versions of the basic arithmetic operators, twice as many weighted connections for each neuron, and activation values treated as pairs of real values. The differences between complex-domain and real-domain neural networks are explained well and illustrated with C++ code.
Chapter 3 discusses some of the most important considerations involved in preparing data for presentation to a neural network. It takes into account the network’s tolerance for noise, nonlinearity, interactions between variables, correlation among inputs, and even total redundancy. The preparation techniques reviewed, discussed, and implemented are used to remove linear trends, seasonal components and slow variation, scaling, and differencing.
Instead of limiting the use of complex-domain neural networks to problems inherently in the complex domain, chapter4 explores the use of these networks to solve problems that treat the data as pairs of values bounded by some semantic relation. This chapter focuses on problems that can use Fourier transforms to detect signals in the frequency domain and process data in the time domain. The discourse follows the same methodology used in the rest of the book: present the concept and techniques and develop an implementation in C++.
Chapter 5 presents functions that map real vectors to matrices of complex numbers that can be interpreted as the description of the time/frequency composition of the input. This is particularly useful in the analysis of time series, but can be generalized to images and other spatiotemporal patterns. The complex-domain neural networks can take the complex time/frequency matrices as input to produce classifications.
Chapter 6 briefly presents some uses of time/frequency localization in conjunction with complex-domain neural networks, namely speech recognition, radar, and phase transition.
Chapter 7 covers the two-dimensional version of the transforms studied in chapter 5. These two-dimensional transforms allow applications beyond signals, including image processing. An important contribution of this chapter is an image processing methodology that guides the process of obtaining and preparing the data, building the neural network model, and interpreting the output.
In chapter 8, shape descriptors based on moments are presented as alternatives to other techniques applied to images in order to segment, focalize, or locate regions that can give a clue to the identity of a whole image. These shape descriptors can be used in the processing of signals (waves) as well as images, and allow the program to obtain second-order features from the image, such as the perimeter, rotation, and precision.
In chapter 9, a different technique is briefly presented, based on the distribution of tones of the pixels in a given area. The value of this approach for satellite and medical imagery is mentioned. The author gives several algorithms to discriminate tones, textures, and contrasts, in order to separate the areas to be presented to the neural network. The code accompanying this chapter is excellent, but the discussion is brief.
Chapter 10 describes the interface with the MLFN program included on the disk that accompanies the book. This chapter is necessary because of the unfriendly program interface, which is based mainly on the processing of specifications written in a command-based language.
The appendix briefly reviews the contents of the disk and the installation process. The bibliography reflects a limited view of the state of the art of signal processing, image processing, and neural networks.
The book is intended to be used as a reference by designers and programmers of signal and image processing applications. From this point of view, it constitutes a valuable resource. The description of each technique and the development of the corresponding algorithm constitute the best feature of the book; as such, I can recommend it to serious developers involved in solving real-world signal and image problems using neural networks.
