How Does Face Detection Actually Work? Part 2
In this post we continue with our previous article on Paul Viola and Michael Jones, the two minds behind the Viola-Jones framework for detecting faces. The Viola-Jones method is the most well known technique for facial recognition, and the real breakthrough for the team came about when they decided not to analyze the image directly.
Viola and Jones decided that they could break up an image into a series of rectangular shapes, or Haar-like features, based their similarity to the observation of detailed waveforms with Haar wavelets. They are simple square-shaped waveforms named after a mathematician from Hungary named Alfred Haar.
Viola and Jones began by converting an original image to a basic greyscale color space, dubbed the YCbCr color space. The Y component represents the intensity/luminance of the original image and the others determine the flesh tones later along in the process. They calculated the pixel’s Y component, or luma, using a formula Y = 0.299R + 0.587G + 0.114B. The R, G, and B are the measurements of red, green, and blue components in the pixel’s value.
Beginning to search for these rectangular, large-scale features, Viola and Jones’ framework scans intensity values among the assortment of rectangular blocks the image is broken up into. The next step was to determine the intensity of each rectangle’s sum, which was defined from a set of intensity values rated from 0 to 255. Using these sums, Viola and Jones could detect darker and lighter rectangular blocks. The adjacent blocks are called ‘features.’
The team categorized a various types of these Haar-like features. Some show one darker block alongside a lighter one, aligned vertically. Another shows the same thing horizontally. Another features a light block placed between two darker rectangles. Another shows a series of 4 blocks, 2 light and 2 dark that are arranged together. Haar-like features may be any scale or at various positions within the image, and each feature’s value is calculated as a sum of pixel intensity in the light rectangle less the sum of pixels present in the darker one.
Next, each feature’s value is filtered through to decide if the feature is in fact there in the image.
Still, this is quite a few quick computations, so Viola and Jones calculated pixel sum of each rectangle from the upper-let corner, or a point where they were able to calculate the integral image, a sum of pixels of the feature. Each point has its own integral image, and they were able to calculate the sum of the pixels overall.
We’ll get to the final stage of the Viola-Jones method in the next article, and again, here’s a link to the original post:
http://www.techradar.com/news/software/applications/how-face-detection-works-703173
Related articles: