Image Matching Based on Trusted Feature in Sketch-based Retrieval

A novel image matching algorithm based on trusted feature in sketch-based retrieval is proposed. On the basis of traditional distance transformation (DT), trusted feature distance is introduced to eliminate those invalid distance values falsely contributed to the similarity calculation between sketch and database images, i.e. non-feature pixels which are actually far away from feature pixels will be ignored in similarity computing. Moreover, image similarity is weighted by matching feature density of the original color image to weaken the interference of a few highly matched pixels on the image similarity. Extensive experiments on various retrieval tasks show better accuracy than traditional DT methods.