engleski

Image Object Recognition Using Color and Spectrum Comparison

Automatic object recognition is important but challenging task in computer vision. The main difficulty is that the system must be able to differentiate specific object in spite of restrictive factors such as lighting conditions due to different object orientations and positions with respect to the camera. In this work, we use a known method when trying to recognize a specific object template as a part of a target image, called template matching. The experiment is implemented as a twostage process. First, images are preprocessed by reducing their resolution, and mean template brightness is stored for normalization during the comparison to eliminate lightness alternations. Furthermore, color-based template matching is applied in order to find areas where object is most likely to be located. Best-matching subparts of a target image are extracted and further analyzed in frequency domain. Our experiment shows that object is successfully recognized on various different scenes while exhibiting various positions with respect to the camera. Moreover, if object has already been recognized during the first step, the second-step analysis can additionally confirm the object recognition. Thus we combine two features, color and spectrum of an object, to exhibit its true location on a target image in a robust way.

object recognition ; template matching ; color space ; spectrum ;

nije evidentirano