There are many people who study face recognition, but what is truly epoch-making is Paul Viola's article "RobustReal-time Object Detection". This article makes
Facial recognition is seeing the light of day in practical applications. why would you said this?
Because, he proposed three revolutionary improvement methods, and applied these methods to actual embedded products, and the effect has been qualitatively improved.
What are these three methods?
The first is an integral graph that can quickly extract image features. The basic idea is to calculate once and use it multiple times. In fact, the integral graph only needs to traverse the image once to find the
The sum of all area pixels greatly improves the efficiency of image feature value calculation.
The second is to select a small number of important features to build cascades through adaboost. The so-called cascade is to connect multiple strong classifiers together to operate. Every strong classifier consists of
It is composed of several weak classifiers weighted. The role of this part is to greatly simplify the model and the amount of computation, provided that the recognition rate does not drop.
The third major contribution is to dynamically increase the detection rate by assembling the complex cascades in a waterfall structure, using a method that intercepts the focal regions of the image.
The author not only has a detailed explanation in theory, but also uses the research results on the handheld device of Compaq iPaq handheld, the main frequency of this device is about 200M, and there is no floating point operation.
This method has very important practical value in embedded devices.