Data compression neural network with winner-take-all function

Claim 2

An image data system, comprising: a lossy image compressor having an image compression ratio in excess of 10 for receiving an original image and producing first compressed image data from said original image, said first compressed image data specifying an address index of a corresponding one of a set of predetermined images; means for computing a difference between said original image and said predetermined image specified by said first compressed image data; a selectively activated lossless image compressor for selectively compressing at least one of said difference and said first compressed image data to produce second compressed image data; wherein said lossy compressor further comprises a neural network being trained to produce states for a respective one of the predetermined images and a winner-take-all circuit for selecting a winning state of said neural network from all possible states of said neural network; and means for transmitting said first and second compressed image data.

Claim 3

The system of claim 1 further comprising: a lossless image de-compressor for receiving said second compressed image data and producing therefrom said difference; a lossy image de-compressor for receiving said first compressed image data and producing therefrom said one predetermined image specified by said first compressed image data; and means for adding said difference to said one predetermined image to produce a reproduction of said original image.

Claim 4

The system of claim 2 wherein said lossy compressor comprises: means for storing said set of predetermined images; wherein said neural network having an input layer and an output layer, said neural network having a plurality of possible states and being trained to produce a respective one of said states at said output layer for a respective one of said predetermined images received at said input layer; wherein said winner-take-all circuit connected to outputs of said output layer for selecting a winning one of said states of said neural network from all possible states of said neural network, wherein said first compressed image data corresponds to said winning state.

Claim 5

The system of claim 3 wherein each of said predetermined images is stored in said means for storing in accordance with a respective image index, said system further comprising: a codebook for correlating a respective image index with a corresponding winning state of said neural network, and wherein said first image data comprises data identifying said winning state.

Claim 6

The system of claim 4 wherein said lossy de-compressor comprises: a second memory for storing said set of predetermined images and for providing one of said predetermined images specified by said first compressed image data.

Claim 7

The system of claim 5 wherein said first compressed image data comprises data identifying said winning state, said second memory stores said predetermined images in accordance with corresponding image indices, said lossy de-compressor further comprising: a codebook correlating each one of said states of said neural network with a corresponding one of said image indices, whereby an image index corresponding to said winning state addresses said second memory to retrieve the corresponding predetermined image.

Claim 8

The system of claim 5 wherein said first memory is updated by replacing the predetermined image corresponding to said winning state with said original image, and said second memory is updated by adding said difference to the corresponding predetermined image in said second memory.

Claim 9

The system of claim 2 wherein said lossless compressor comprises: plural entropy encoders each employing a different entropy code; an optimum code selector for computing a code length of successive blocks of said second image data and choosing for each said block one of said plural entropy encoders whose code is optimum for the code length of that block; and means for generating data identifying the code chosen by said optimum code selector for transmission with said second compressed image data.

Claim 10

The system of claim 8 wherein said lossless de-compressor comprises: plural entropy decoders; and means responsive to said data identifying the code chosen by said optimum code selector of said lossless compressor for choosing a corresponding one of said plural entropy decoders to decode said second image data.

Claim 11

The system of claim 1 wherein said lossy compressor comprises: a first memory for storing said set of predetermined images; wherein said neural network having an input layer and an output layer, said neural network having a plurality of possible states and being trained to produce a respective one of said states at said output layer for a respective one of said predetermined images received at said input layer; wherein said winner-take-all circuit connected to outputs of said output layer for selecting a winning one of said states of said neural network from all possible states of said neural network, wherein said first compressed image data corresponds to said winning state.

Claim 12

The system of claim 10 wherein each of said predetermined images is stored in said first memory in accordance with a respective image index, said system further comprising: a codebook for correlating a respective image index with a corresponding winning state of said neural network, and wherein said first image data comprises data identifying said winning state.

Claim 13

The system of claim 10 wherein said first memory is updated by replacing the predetermined image corresponding to said winning state with said original image, whereby said lossy compressor is adaptive to images from new scenes.

Claim 14

The system of claim 10 wherein said input and output layers are linked by an array of synapses, said system further including a learning processor comprising: means responsive to identification of said winning state by said winner-take-all circuit for updating synapses associated with said winning state in accordance with the corresponding one of said predetermined images applied to said input layer during a training exercise; means for storing the image index of the predetermined image applied to said input layer during a training exercise and an index identifying the corresponding winning state as a codebook entry.

Claim 15

The system of claim 1 further comprising a lossy compressor for compressing said first image data.

Claim 16

The system of claim 1, further comprising: a bypass device for routing said first compressed image data to bypass said lossless image compressor so that said lossless image compressor does not compress said original image.

Claim 17

The system of claim 15, further comprising a switching device for selectively activating and deactivating said bypass device.

Claim 18

The system of claim 1, wherein said neural network is adaptable and is trained in real time.

Claim 19

An image data system, comprising: a lossy image compressor having an image compression ratio in excess of 10 for receiving an original image and producing first compressed image data from said original image, said first compressed image data specifying an address index of a corresponding one of a set of predetermined images; means for computing a difference between said original image and said predetermined image specified by said first compressed image data; a lossless image compressor for selectively compressing at least one of said difference and said first compressed image data to produce second compressed image data; bypass means for routing said first compressed image data to bypass said lossless image compressor so that said lossless image compressor does not compress said original image first compressed image data; wherein said lossy compressor further comprises a neural network being trained to produce states for a respective one of the predetermined images and a winner-take-all circuit for selecting a winning state of said neural network from all possible states of said neural network; and means for transmitting said first and second compressed image data.

Claim 20

An image data system, comprising: a lossy image compressor having an image compression ratio in excess of 10 for receiving an original image and producing first compressed image data from said original image, said first compressed image data specifying an address index of a corresponding one of a set of predetermined images; means for computing a difference between said original image and said predetermined image specified by said first compressed image data; a lossless image compressor for selectively compressing at least one of said difference and said first compressed image data to produce second compressed image data; bypass means for routing said first compressed image data to bypass said lossless image compressor so that said lossless image compressor does not compress said original image first compressed image data; wherein said lossy compressor further comprises a neural network being trained to produce states for a respective one of the predetermined images and a winner-take-all circuit for selecting a winning state of said neural network from all possible states of said neural network; and means for transmitting said first and second compressed image data.

Claim 21

The system of claim 18, further comprising means for selectively activating and deactivating said bypass means.

Claim 22

The system of claim 19 wherein said lossy compressor comprises: a first memory for storing said set of predetermined images; a neural network having an input layer and an output layer, said neural network having a plurality of possible states and being trained to produce a respective one of said states at said output layer for a respective one of said predetermined images received at said input layer; a winner-take-all circuit connected to outputs of said output layer for selecting a winning one of said states of said neural network from all possible states of said neural network, wherein said first compressed image data corresponds to said winning state.

Claim 23

The system of claim 20 wherein each of said predetermined images is stored in said first memory in accordance with a respective image index, said system further comprising: a codebook for correlating a respective image index with a corresponding winning state of said neural network, and wherein said first image data comprises data identifying said winning state.

Claim 24

The system of claim 18, wherein said neural network is adaptable and is trained in real time.