Haar Features

Haar features (the name derives from their resemblance to Haar wavelets) are a set of image filters formed by summing and subtracting purely rectangular subregions of the image itself (PP99). Examples of Haar features are shown in Figure 7.1. The resulting filter value is the sum of the gray levels of the pixels covered by the white areas, minus the value of the pixels covered by the areas shown in black. By their nature, these filters can be implemented efficiently using the integral image (Section 1.15).

Haar features are used as approximations of convolutions to compute feature points in the SURF algorithm, or as input features for decision trees to obtain weak classifiers.

Figure 7.1: Examples of Haar features. In the light and dark areas, the covered area is respectively added or subtracted.

Although the shape could potentially be arbitrary, the number of bases for the features is normally limited (complex and computationally expensive features are avoided whenever possible).

In addition to the type of feature, the subregion to which it is applied must be selected: from each subwindow of the area being analyzed, a value can be extracted by applying one of these many features. Determining which features are most discriminative is part of the training process (Decision Stumps ordered using AdaBoost) or can be done using techniques such as PCA.