# US 12,217,467 B2 · System and method for image compression based on machine learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,217,467 B2
- Document ID: US12217467B2
- Status: Granted
- Kind: B2
- Application number: 17/571,538
- Filed: 2022-01-10
- Published: 2022-07-14
- Granted: 2025-02-04
- Earliest claimed: 2021-01-08
- Applicants: Averroes.ai Inc.
- Inventors: Omar Ahmad Abdo Al Tamimi; Tareq Aljaber
- Web page: https://patentgrep.com/patent/US12217467B2

## CPC

- G06N 3/08: Learning methods
- G06T 9/002: {using neural networks}
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/09: Supervised learning
- G06T 3/4046: using neural networks
- H04N 19/124: Quantisation
- H04N 19/147: according to rate distortion criteria (rate-distortion as a criterion for motion estimation H04N19/567)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- H04N: PICTORIAL COMMUNICATION, e.g. TELEVISION

## Abstract

A system and method for compressing and decompressing image data, which provides better compression and minimal representation of the input image with minimal loss compared to previous Deep Learning codecs. The system can provide for quantization during training, flexible addition of filters, and conditional complexity of image compression. The Deep Learning codec generates codes that are directly usable with Machine Learning algorithms, thus boosting the performance of Machine Learning algorithms.

## Claims

### Claim 1

A method for dynamic compression of image data with high compression ratio and high fidelity, the method implemented in a system comprising a processor and a memory, the method comprising the steps of:
constructing a convolution neural network based quantized autoencoder network comprising an encoder network, a bottleneck network, and a decoder network,
wherein the encoder network comprises one or more convolution compression blocks each having two convolution layers, the encoder network upon execution by the processor receives an input image data and reduces dimensions of the input image data by dynamic processing the input image data using the one or more convolution compression blocks in series,
wherein the bottleneck network upon execution by the processor receives an unquantized compression representation of the input image data as an output of the encoder network, wherein the bottleneck network comprises fake quantization module, a secondary encoder, and a secondary decoder,
wherein the decoder network upon execution by the processor receives a quantized compressed representation of the input image data as an output of the bottleneck network, restructure the quantized compressed representation of the input image data to obtain a compressed output image, the decoder network comprises a plurality of forward blocks.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the one or more convolution compression block comprises a Conv2D filter and an exponential linear unit (ELU) filter connecting in a series.

### Claim 3 (depends on 1)

The method according to claim 1, wherein the encoder network is configured to grow in layers by adding further one or more of the convolution compression blocks in series, wherein a compression ratio increases fourfold for each growth stage.

### Claim 4 (depends on 3)

The method according to claim 3, wherein the encoder network has a compression ratio of (16×2n):1, where n is a number of the one or more convolution compression blocks in the encoder network.

### Claim 5 (depends on 4)

The method according to claim 4, wherein the encoder network is trained by supervised Greedy layer-wise Pretraining to dynamically derive a value of n.

### Claim 6 (depends on 5)

The method according to claim 5, wherein the method further comprises the steps of:
upscaling the quantized compressed representation of the input image data, by the decoder network, upon execution by the processor, for n number of times, by subjecting to n layers of UpSampling2D layer.

### Claim 7 (depends on 1)

The method according to claim 1, wherein the secondary encoder is a JPEG encoder, and the second decoder is a JPEG decoder.

### Claim 8

A system for dynamic compression of image data with high compression ratio and high fidelity, the system comprising a processor and a memory, the system configured to implement a method comprising the steps of:
constructing a convolution neural network based quantized autoencoder network comprising an encoder network, a bottleneck network, and a decoder network,
wherein the encoder network comprises one or more convolution compression blocks each having two convolution layers, the encoder network upon execution by the processor receives an input image data and reduces dimensions of the input image data by dynamic processing the input image data using the one or more convolution compression blocks in series,
wherein the bottleneck network upon execution by the processor receives an unquantized compression representation of the input image data as an output of the encoder network, wherein the bottleneck network comprises fake quantization module, a secondary encoder, and a secondary decoder,
wherein the decoder network upon execution by the processor receives a quantized compressed representation of the input image data as an output of the bottleneck network, restructure the quantized compressed representation of the input image data to obtain a compressed output image, the decoder network comprises a plurality of forward blocks.

### Claim 9 (depends on 8)

The system according to claim 8, wherein the one or more convolution compression block comprises a Conv2D filter and an exponential linear unit (ELU) filter connecting in a series.

### Claim 10 (depends on 8)

The system according to claim 8, wherein the encoder network is configured to grow in layers by adding further one or more of the convolution compression blocks in series, wherein a compression ratio increases fourfold for each growth stage.

### Claim 11 (depends on 10)

The system according to claim 10, wherein the encoder network has a compression ratio of (16×2n):1, where n is a number of the one or more convolution compression blocks in the encoder network.

### Claim 12 (depends on 11)

The system according to claim 11, wherein the encoder network is trained by supervised Greedy layer-wise Pretraining to dynamically derive a value of n.

### Claim 13 (depends on 12)

The system according to claim 12, wherein the method further comprises the steps of:
upscaling the quantized compressed representation of the input image data, by the decoder network, upon execution by the processor, for n number of times, by subjecting to n layers of UpSampling2D layer.

### Claim 14 (depends on 8)

The system according to claim 8, wherein the secondary encoder is a JPEG encoder, and the second decoder is a JPEG decoder.