| Summary: | 碩士 === 國立成功大學 === 電機工程研究所 === 82 === In this thesis,a new image compression method is devised.
Discrete Cosine Transform is the kernel of the compressor. The
new technique classifies the image subblock into eights classes
by their characteristics which can be got from the
transfomed domain.Four edge classes,three texture classes ,one
smooth class.This kind of classification is called Three Model
Classification.As to the overhead of classification ,finite
state concept is used to reduce the overhead by predicting the
current block's class from the previously blocks. In order to
promote the correct ratio of prediction,edge orientation should
be considered.We exploit the relationship between the
transformed domain and edge orientation.As we know, the smooth
regions occupy most part of a natual images.Larger coding size
can get higher compression ratio in the smooth regions. But
this will sacrifice the quality of complicated regions.In order
to solve this deficiency.We devised a variable block coding
size algorithm.The edge blocks use the fixed 8*8 block to keep
the detailed parts and the variable block size segmentation
scheme is applied to texture and smooth regions. The new
segmentation method is called "class driven segmentation" The
overhead of the segmention is zero. The simulation results show
good quality for the decoded images.
|
|---|
