Fast Template Matching Methods And Their Applications In Eyebrow Recognition

Template matching is a common task in machine vision, image and video processing, i.e., how to find a template from one given image quickly and accurately, and its related methods have been widely used in manufacturing quality control, texture synthesis, motion estimation, object detection, road tracking, behavior recognition and image denoising etc. However, it is still a challenge task to design better fast template matching methods. So, this paper studied fast template matching methods systematically and deeply, and then apply them to eyebrow recognition. Eyebrow recognition has the basic idea that human eyebrow can be seen as an independent biometric, and used to determine the identity of an individual with a high probability of success. It can be considered, and replace iris recognition when the iris is difficult for collecting, or replace face recognition when the portions of the face following the noses were occluded. But existing eyebrow recognition methods all had the problems of the manual processing, and low recognition rate. Applying fast template matching methods to eyebrow recognition can be conducive to design auto recognition methods, and improve the recognition rate. The main innovative achievements of this paper are as follows:(1) Propose a fast orthogonal Haar transform template matching algorithmAlthough the orthogonal Haar transform template matching algorithm(OHT) may have good performance by using images strip sum, but it still needs 3 calculations for each Haar projection vale. By establishing a solid mathematical model for orthogonal Haar transform, this paper based on the concept of square sum, proposes a fast orthogonal Haar transform template matching algorithm(FOHT). It reduces the calculation of per Haar projection vale from 3 subtractions to 1, so that it can obtain higher speed. The experimental results showed that FOHT is very competitive with OHT in most cases of matching one single template; and general faster than OHT in all cases of matching multi templates.(2) Propose a quasi Haar transformation template matching algorithmSince OHT and FOHT require that templates must be a standard size, but they are always arbitrary sizes in fact. So it is necessary to develop quasi Haar transform from orthogonal Haar transform. By establishing a solid mathematical model for quasi Haar transform, this paper based on tree division strategy, which makes the calculations of per quasi Haar projection value reducing to almost half of that direct computing, proposes a quasi Haar transformation template matching algorithm(QHT). A large number of experiments showed that, QHT can directly deal with arbitrary size of templates with a high efficiency; and it would faster than OHT in some cases of matching standard size of templates.(3) Propose a novel idea of matching-recognizing frameworkUnder the traditional detecting-recognizing framework, eyebrow recognition methods require manual processing, which has brought uncertainty factors. So, this paper proposed a novel idea of matching-recognizing framework, which firstly matched the most similar regions of templates, and then identified the results by a discriminative distance. Experiments showed that the framework can establish the automated system of eyebrow recognition. And it also demonstrated the feasibility of replacing face recognition by eyebrow recognition. In addition, matching-recognizing framework is a generic framework of automatic image recognition, and it can be used for other image biometrics with only replacing eyebrow templates.(4) Propose 4 new eyebrow recognition methods based on fast template matchingIn order to improve the eyebrow recognition accuracy and efficiency further, the OHT, FOHT, QHT, and gradient template matching algorithm(GTM) were applied to the eyebrow recognition respectively, and 4 new eyebrow recognition methods were obtained. Experiments show that these eyebrow recognition methods have high efficiency. And, in the publicly available eyebrow database BJUTED, the method using GTM achieved 98.12% recognition rate, which is the highest recognition rate for BJUTED so far.