Chap 1 Pattern Analysis and Statistical Learning

Chapter 1 sets out the basics of how a pattern recognition system will operate. This discusses the key elements that systems need and how they are related to image processing/classification. Classical statistical classification methods are mentioned and pitfalls such as the 'curse of dimensionality' are highlighted.  This introduction that sets the tone for the book. The means to teach and educate a classification type system are covered, which is often overlooked in books.  This gives a gentle introduction to readers of various levels of knowledge.

Chap 2 Unsupervised Learning

Unsupervised learning gives an indication as to how the reader can automatically find the classes and distributions needed for classification to take place. The authors show the different techniques that can be used depending on the number of classes you are using. They cover various techniques including Gaussian Mixtures for modelling the underlying class probabilities whose clusters are discovered through some form of unsupervised learning. Included in this chapter is the ever interesting clustering based on the Gestalt principle. This principle is based on the human visual system and how it perceives what are commonly known as optical illusions, or tricks on the eye.

Chap 3 Component Analysis

Often we find in classification and analysis tasks that the number of dimensions of the space that we are working in is large, and directly effects performance of our system. Zheng and Xue take us through the practice of Component Analysis to reduce this dimensionality while retaining information. Here, the usual suspects of Independent Component Analysis, Principle Component Analysis, etc. are found. This is a key feature of systems, and it is good to see a detailed description of the method and its implementation, pointing out to the reader problems that can occur and how they can be overcome.

Chap 4 Manifold Learning.

Modern advances in dimension reduction have lead to the domain of Manifold Learning, it is good that this has been included. The authors take care in not only describing the method but also in providing an introduction to the necessary supporting methods and maths. This provides the reader with the tools that they need to build up an understanding of Manifold Learning. Care was taken to remember that the book will be read by people of varying levels of understanding and, as such, that some of the supporting work is included. This can accelerate a novice reader’s rate of learning.

Chap 5 Functional Approximation

Wavelet Transforms and filters are a useful tool for image compression. Importantly the authors discuss the usefulness of  ‘motion compensation’ for use within video codecs which is soon to be a part of the SVC standards.

Chap 6 Supervised Learning for Visual Pattern Recognition

When the training data have class labels applied, the learning becomes a question of how best to model that class representation, looking at distributions and class boundaries. This is generalised to a classification function. The authors discuss how this can be attained using a variety of learning methods but focus is provided on the popular support vector machine and boosting algorithm. These remain active areas of useful research and the authors do well by adding more depth of discussion on these two techniques. I feel that there are some other even more basic supervised learning techniques that could be included to keep with the theme of the book to include material for the more novice readers.

Chap 7 Statistical Motion Analysis

Motion analysis is necessary for a variety of image and video processing tasks. It allows for a better understanding of what is occurring within the images over time. Zheng and Xue bring to our attention the concepts of optical flow and the various methods available for calculation. They describe well how this can be used at the pixel level or at a higher level with model based motion analysis. This is then extended to cover how motion based segmentation is used to extract moving objects from a sequence of images.

Chap 8 Bayesian Tracking of Visual Objects

Tracking of objects in Computer Vision is needed for many real world applications. There are some very well used and robust trackers available that are tried and tested. So long as you understand your problem well you should be able to track within its constraints. Zheng and Xue provide the standard approaches which many are familiar with such as the Kalman, Monte Carlo (Particle Filters). These filters are well described providing enough information to implement and work with these. It also serves as a good source of reference for those more familiar with these topics. I think it should be noted that these techniques are heavily covered in other texts as well, and although an important technology, nothing new is given by this description.

Chap 9 Probabilistic Data Fusion for Robust Visual Tracking

The only chapter that approaches fusion methods is for Visual Tracking schemes. This looks at how multiple trackers are used to fuse at different levels within a tracking system that receives data from a number of sensors. This is a very real world problem and this chapter attempts, along with the following multi target trackers, to shed some light on what is a highly complex problem. The level of prior knowledge needed is quite high, but it would be naïve to expect anything less. A whole book could be written on this subject. All of the facts are laid out in a thoughtful and well explained manner for the more experienced reader.

Chap 10 Multitarget Tracking in Video-Part I

The human perception and visual system is looked at in the final chapter. This discusses how we can learn from ourselves and better educate computational methods. AI has long been an area of research and has a huge impact on the processing of images by computers. This is something that I find interesting and  think that more reference can be made to this with respect to other techniques covered in the book. It gives a slightly different angle and can help novices get a better understanding, or capture their imagination a little better.

 

Overall I felt that a very broad subject area was covered by the book. In some chapters it was obvious that the level for which the text was aimed was quite introductory, giving a well executed explanation of not just the technique, but also the supporting techniques. This would serve the book well as a tool to someone learning the technique from new, such as material to support a taught course. But other chapters were definitely aimed at a more educated reader, this maybe puts the book at cross purposes? Is it covering all of the basics for newcomers, or is it looking at advances in technology, or both? Due to the nature of the subject area there needs to be complex and technical chapters. There is a lot of information that is placed within a single book, which I think it does quite well.

Personally I would prefer to see some more assistance with the setting up and implementation of the areas covered by the book, this would be really helpful for newcomers. Also some more discussion on fusion and how vision/AI techniques are related to our own visual perception systems. These are areas where image/video processing will move to in the future as we try to gain a better understanding of the content. This could provide a more interesting read and make it less of a reference book, which it does appear to be trying to do.

Overall I enjoyed the book, and learned some things I didn't know, but should have! I found that the subjects were well discussed and at a level that suited my knowledge. I would recommend it as a general purpose book for image and video analysis, but would probably then look for something more focussed if I were working on a particular area.