mobile-menu

Perceptual Compression

Perceptual compression is a lossy compression technique for video and audio data. Unlike other lossless and lossy algorithms based on mathematics and logic, perceptual compression is based on the temporal observation of the capabilities of human sensory systems.

The idea is to compress a stream of data by omitting valid but unimportant data that a human viewer/listener might not observe or consider important. A very weak analogy would be the compression of the text, skipping non-critical vowels.

MP3 compression of audio tries to reject sounds that cannot be heard well or not at all because they are closer in frequency to louder sounds or follow a louder sound closely over time. It also reproduces louder sounds with greater precision than smaller ones. All these measures allow less information to be recorded in the recorded/transmitted data and thus result in fewer bits in the data. Traditional lossless compression is also added to MP3 after perceptual compression.

Perceptual compression of video typically encodes more information for luminescence ("luminance") than color, as the human eye is more sensitive to changes in brightness than color changes. Other features of human vision are also utilized in different ways with different compression schemes.

Since perceptual compression is ad hoc, it's possible that each compressor will produce somewhat different compressed data. In the case of MP3, the specification does not define compression, only what a compressed file should look like for a compressor to work. Presumably this will apply to most perceptual compression schemes.

Different compression schemes may work better with different materials, and all may have issues in certain situations. Other problems may intervene, such as difficulties in recompressing data that was previously compressed and then decoded, or difficulties in retrieving compressed data in the middle of the stream.

Perceptual compression is used together with audio and video data. It will likely be used in the future along with other sensory data types.

 

Latest Updates on Feb 14, 2022