Cascading classifiers

Cascading is a particular case of ensemble learning based on the concatenation of several classifiers, using all information collected from the output from a given classifier as additional information for the next classifier in the cascade. Unlike voting or stacking ensembles, which are multiexpert systems, cascading is a multistage one. Cascading classifiers are trained with several hundred "positive" sample views of a particular object and arbitrary "negative" images of the same size. After the classifier is trained it can be applied to a region of an image and detect the object in question. To search for the object in the entire frame, the search window can be moved across the image and check every location with the classifier. This process is most commonly used in image processing for object detection and tracking, primarily facial detection and recognition. The first cascading classifier was the face detector of Viola and Jones (2001). The requirement for this classifier was to be fast in order to be implemented on low-power CPUs, such as cameras and phones.

Cascading is a particular case of ensemble learning based on the concatenation of several classifiers, using all information collected from the output from a given classifier as additional information for the next classifier in the cascade. Unlike voting or stacking ensembles, which are multiexpert systems, cascading is a multistage one. Cascading classifiers are trained with several hundred "positive" sample views of a particular object and arbitrary "negative" images of the same size. After the classifier is trained it can be applied to a region of an image and detect the object in question. To search for the object in the entire frame, the search window can be moved across the image and check every location with the classifier. This process is most commonly used in image processing for object detection and tracking, primarily facial detection and recognition. The first cascading classifier was the face detector of Viola and Jones (2001). The requirement for this classifier was to be fast in order to be implemented on low-power CPUs, such as cameras and phones. Each stage must validate all faces, and can produce many false positives. For example, if stage 1 were to mark as 'does not contain a face' 20% of rectangles containing a face (false negative rate=20%), then the total performance of the chain cannot be higher than 80% true positive, whatever the next stages are, since 20% of faces have already been rejected. This suggests that a good stage must have 100% true positive and for example 40% false positive, that is accept all rectangles containing faces and erroneously mark many rectangles as potentially containing a face, to be eliminated by later stages. For a first stage, 100% true positive and 40% false positive still gives a lot of false negative, if only 1 in a 1000 rectangles in an image contain a face, there will still be 400 to 1 false possible faces after the first stage. If the first stage is very fast (a few operations), we have eliminated 60% of rectangles not containing a face very quickly. The training procedure for one stage is therefore to have many weak learners (simple pixel difference operators), train them as a group (raise their weight if they give correct result), but be mindful of having only a few active weak learners so the computation time remains low. The first detector of Viola & Jones had 38 stages, with 1 feature in the first stage, then 10, 25, 25, 50 in the next five stages, for a total of 6000 features. The first stages remove unwanted rectangles rapidly to avoid paying the computational costs of the next stages, so that computational time is spent analyzing deeply the part of the image that have a high probability of containing the object.

Sources

Text derived from https://en.wikipedia.org/w/index.php?oldid=1326978020 under CC BY-SA 4.0; modifications have been made.