Self-adaptive neural net based vector quantizer for image compression

Abstract

Abstract A self adaptive neural net based vector quantizer for image compression uses a conventionally designed small front-line code book and a larger second-line code book of vectors which quantize the space finely. Only the front-line code book is used for encoding. The front-line and second-line vectors are connected in a neural net, their strength of interconnection dictated by their correlation. Vectors are exchanged between code books on a periodic basis. The movement of a vector is dictated by a rising function for second-line vectors and a falling function for front-line vectors, which functions are computed from interconnections weights, average error and frequency of usage. After a periodic exchange of vectors, the affected nodes on the neural net are recomputed. Thus, the front-line code book is periodically revised to adapt to the images being coded.

Inventors

• 👤 Sudhir S Dixit
• 👤 Yushu Feng

Bibliographic Details