Global Methods
Most of the early approaches to object recognition rely on a global object model. In this context “global” means that the model represents the object to be recognized as a whole, e.g., by one data set containing several global characteristics of the objec
- PDF / 1,500,232 Bytes
- 29 Pages / 439.37 x 666.142 pts Page_size
- 21 Downloads / 216 Views
Global Methods
Abstract Most of the early approaches to object recognition rely on a global object model. In this context “global” means that the model represents the object to be recognized as a whole, e.g., by one data set containing several global characteristics of the object like area, perimeter, and so on. Some typical algorithms sharing this object representation are presented in this chapter. A straightforward approach is to use an example image of the model to be recognized (also called template) and to detect the object by correlating the content of a scene image with the template. Due to its simplicity, such a proceeding is easy to implement, but unfortunately also has several drawbacks. Over the years many variations aiming at overcoming these limitations have been proposed and some of them are also presented. Another possibility to perform global object recognition is to derive a set of global features from the raw intensity image first (e.g., moments of different order) and to evaluate scene images by comparing their feature vector to the one of the model. Finally, the principal component analysis is presented as a way of explicitly considering expected variations of the object to be recognized in its model: this is promising because individual instances of the same object class can differ in size, brightness/color, etc., which can lead to a reduced similarity value if comparison is performed with only one template.
2.1 2D Correlation 2.1.1 Basic Approach 2.1.1.1 Main Idea Perhaps the most straightforward approach to object recognition is 2D cross correlation of a scene image with a prototype representation of the object to be found. Here, the model consists of a so-called template image, which is a prototype representation of the gray value appearance of the object. Model generation is done in a training phase prior to the recognition process. For example, the template image is set to a reference image of the object to be found. 2D correlation is an M. Treiber, An Introduction to Object Recognition, Advances in Pattern Recognition, C Springer-Verlag London Limited 2010 DOI 10.1007/978-1-84996-235-3_2,
11
12
2 Global Methods
example of an appearance-based scheme, as the model exclusively depends on the (intensity) appearance of the area covered by the “prototype object” in the training image. The recognition task is then to find the accurate position of the object in a scene image as well as to decide whether the scene image contains an instance of the model at all. This can be achieved with the help of evaluating a 2D cross correlation function: the template is moved pixel by pixel to every possible position in the scene image and a normalized cross correlation (NCC) coefficient ρ representing the degree of similarity between the image intensities (gray values) is calculated at each position: H W IS (x + a, y + b) − IS · IT (x, y) − IT
ρ(a, b) =
x=0 y=0 H H W W 2 2 IS (x + a, y + b) − IS · IT (x, y) − IT x=0 y=0
(2.1)
x=0 y=0
where ρ(a, b) is the normalized c
Data Loading...
