Visual Object Recognition using Template Matching
Luke Cole, David E. Austin, Lance Cole · 2004
Template matching is well known to be an expensive operation when classifying against large sets of images as the execution time is proportional to the size of the set. This paper addresses the scaling problem by reduction of the size of the set against which new images are compared. A two-tier hierarchical representation of images for each object is created, reducing the set of images to match against. The reduced set is generated via a pre-processing task of extracting areas of interest (blobs) and then reducing the number of blobs found by clustering. The results are based on a raw database of 90 classes and a total of 140, 000 images each of size 680x480. From the 91 classes, over 100, 000 blobs were extracted as the training set which then was reduced by over 90%. For a recognition rate of 90%, only 0.59 % of this training set was examined, giving us an execution time to identify a new image in 6.7 seconds. 1