Compression for recognition and content-based retrieval

Most compression algorithms developed to date aim at achieving the best perceptual quality of the decoded media for the given rate. In this paper we consider several scenarios where the end user of the compressed data is not a human viewer or listener, but rather a known classifier or recognizer. Drawing from applications in speech recognition and image classification, as well as from simple examples, we discuss the new requirements that are imposed on the encoders under these circumstances. Our goal is to motivate the importance, and describe the associated design challenges, of achieving compression optimized for classification/recognition, rather than perceptual quality.