Learning from Examples

Nowadays, machine learning is very often used to solve difficult machine vision problems. In fact, it is a rich field of research encompassing many important concepts that would deserve an entire cookbook by themselves. This chapter surveys some of the main machine learning techniques and explains how these can be deployed in computer vision systems using OpenCV.

At the core of machine learning is the development of computer systems that can learn, by themselves, how to react to data inputs. Instead of being explicitly programmed, machine learning systems automatically adapt and evolve when examples of desired behaviors are presented to them. Once a successful training phase is completed, it is expected that the trained system will output the correct response to new unseen queries.

Machine learning can solve many types of problems; however, our focus here...

Recognizing faces using the nearest neighbors of local binary patterns

Our first exploration of machine learning techniques will start with what is probably the simplest approach, namely, nearest neighbor classification. We will also present the local binary pattern (LBP) feature, which is a popular representation encoding the textural patterns and contours of an image in a contrasting and unique way.

Our illustrative example will concern the face recognition problem. This is a very challenging problem that has been the object of numerous researches over the past 20 years. The basic solution we present here is one of the face recognition methods that is implemented in OpenCV. You will quickly realize that this solution is not very robust and works only under very favorable conditions. Nevertheless, this approach constitutes an excellent introduction to machine learning and to...

Finding objects and faces with a cascade of Haar features

In the previous recipe, we learned about some of the basic concepts of machine learning. We demonstrated how a classifier can be built by collecting samples of the different classes of interest. However, for the approach that was considered in this previous recipe, training a classifier simply consists of storing all the samples' representations. From there, the label of any new instance can be predicted by looking at the closest (nearest neighbor) labeled point. For most machine learning methods, training is a relatively iterative process, during which machinery is built by looping over the samples. The performance of the classifier produced gradually improves as more samples are presented. Learning eventually stops when a certain performance criterion is reached, or when no more improvements can be obtained from...

Detecting objects and people using SVMs and histograms of oriented gradients

This recipe presents another machine learning method, the SVM, which can produce accurate 2-class classifiers from training data. They have been largely used to solve many computer vision problems. This time, classification is solved by using a mathematical formula that looks at the geometry of the problem in high-dimension spaces.

In addition, we will also present a new image representation that is often used in conjunction with SVMs to produce robust object detectors.

Getting ready

Images of objects are mainly characterized by their shape and textual content. This is the aspect that is captured by the Histogram of Oriented Gradients (HOG) representation...