uses contextual attributes of the digital object to interpret the artifact in a manner that more closely corresponds with human perceptual categories. For example, perceptual hashes allow the matching of visually similar images and are unconcerned with the low-level details of how the images are persistently stored. Semantic methods tend to provide the most specific results but also tend to be the most computationally expensive ones.
Source(s):
NIST SP 800-168