Facsimile and static presentation processing – Facsimile – Auxiliary signal
Patent
1991-07-25
1994-11-29
Brinich, Stephen
Facsimile and static presentation processing
Facsimile
Auxiliary signal
358430, H04N 140
Patent
active
053695038
DESCRIPTION:
BRIEF SUMMARY
The present invention relates to a method of block-compression of digital pictures by autoorganisation of a neural network.
The technique of picture compression consists in reducing the quantity of data required to represent a picture. This technique is widely used in the fields of the transmission or storage of pictures.
When a compressed picture is decoded in order to reconstruct its original form, the reconstructed picture is affected by distortion. Consequently, the efficiency of a compression algorithm is measured simultaneously by its rate of compression, the resulting distortion and its complexity of implementation. The complexity of implementation is a particularly important point in the hardware construction of coders and decoders.
The main known methods of block-compression of digital pictures use algorithms such as the Karhunen-Loeve transformation, the Hadamard transformation, the cosine transformation and vector quantification. These algorithms are implemented on sequential machines for which the data processing is slow and manual manipulations are required when having to adapt to a different picture type.
In artificial intelligence, neural-connective or neuro-imitative networks are presently experiencing a revival of interest, and the architecture of neural networks possesses intrinsically a massive parallel processing potential.
Parallel data processing permits a considerable time saving as against sequential machines, but the algorithms used hitherto for picture compression are adapted to sequential machines and parallelisation of them is not easy.
The invention relates to a new method of picture compression which is completely automatic whatever the picture type to be compressed, is simple to employ, and uses an algorithm which is directly compatible with the future massively parallel connective machines which will be developed in the coming years. This method is based on the learning concept, which enables it to best adapt to a particular picture type, such as television pictures, aerial pictures, etc., and hence to provide low distortion for this picture type. Moreover, the system can be rapidly reconfigured in order to adapt to another picture type simply by recommencing learning on pictures of this type.
According to the invention, the method of picture compression by auto-organisation of a neural network is characterised in that it comprises:
a preprocessing during which pictures of identical type are partitioned into pixel blocks with dimensions smaller than those of the pictures,
a phase of initialising a neural network comprising N neurons disposed in a single plane and having a number of inputs equal to the dimension of the blocks extracted from the pictures, during which values of coefficients, or weights, are randomly allocated to each connection linking a neuron j to an input ei, and an initial neighbourhood radius Vo determining the neighbour neurons of the neuron j is chosen
a learning phase during which the neural network is fed successively with pixel blocks extracted randomly from the various pictures and the weights of the neural connections evolve by successive approximations following an algorithm simulating the behaviour of the visual cortex, through a computation of potentials which aims to elect one neuron with updating of its neighbours, until their values are as close as possible to most probable input values,
a compression phase during which the source picture is partitioned into pixel blocks, each block being presented one after the other to the inputs of the neural network and coded as a function of the neuron of which the values of the weights are closest to the inputs.
Other features and advantages of the invention will emerge clearly from the following description given by way of non-limiting example and made with reference to the attached figures which show:
FIG. 1, the general structure of Kohonen neural networks;
FIG. 2, an example of a neural network according to the invention;
FIG. 3, a flow diagram of the learning phase, according to the invention;
FIG.
REFERENCES:
patent: 5041916 (1991-08-01), Yoshida et al.
Burel Gilles
Pottier Isabelle
"Thomson-CSF"
Brinich Stephen
LandOfFree
Method of picture compression by auto-organisation of a neural n does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method of picture compression by auto-organisation of a neural n, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of picture compression by auto-organisation of a neural n will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-77225