Classification of Building Images in Video Sequences

2. Edge-Based Building Image Classification

Edge-based building image classification can be found in (Iqbal, 1999), (Vailaya, 2001), (Wang, 2002).

Let x be an image or video frame, f = {f(1), . . ., f(n)} be feature sets where f is a function of the image x, E^_t = (e_1^t , . . ., e_5^t ) be a pattern extracted from a feature vector f^(t), and Y = {B, NB} be a class set (B:Building, NB:Non Building). Edge-based building image classification uses a function:g(x): E^t → Y (1) where E^t is extracted from the description of x given by the MPEG-7 edge histogram descriptor.

This descriptor uses an 80-bin histogram to represent the local distribution of directional (vertical, horizontal, diagonal 45^◦, and diagonal 135^◦) and non- directional (isotropic) edges.

g(x) is approximated with a set R = {R₁, . . ., R_C } of if-then fuzzy rules which has the following structure:^WR_k: If e_i^t is A_i^k then Y is y_j(2) where ^w R_k is a weighted rule, w ∈ [0, 1] is the weight, e_i^t a specific type of edge, and A^k is a linguistic label used to transform values from a continuous to a discrete domain.

The fuzzy model has five input variables because of the semantics of edge histogram descriptor which uses five types of edges. Besides, a group of three fuzzy sets is associated with each variable. Thus, rule base is a multi-input single-output space as is depicted in Figure 1.

Figure 1.

Fuzzy model for edge histogram descriptor

R_i and R_j are if-then rules associated with classes NB and B, respectively. H, M, L stand for High, Low, Medium, and Low.