Artificial Intelligence-Driven Image Data Compression and Decompression System and Method

Abstract

This invention relates to an AI-driven image data compression and decompression system and method, comprising a first device and a second device. The first device segments, encodes, quantizes, compresses, and encrypts the image data, then transmits the encrypted image data to the second device. The second device decrypts, decompresses, decodes, and combines the encrypted image data to generate reconstructed image data. This invention can select higher quality or higher compression. Through a trade-off between compression ratio and distortion, this invention can significantly reduce data volume and lower bandwidth requirements for data transmission (e.g., for 5G or 6G transmission), while also reducing storage space (e.g., network, cloud data center, or local device storage) and maintaining the quality of the reconstructed image.

Claims

1. An AI-driven image data compression and decompression system, comprising: a first device, including: a database; an image selection unit, signal-connected to the database; an image segmentation unit, signal-connected to the image selection unit; a neural network coding unit, signal-connected to the image segmentation unit; a quantization compression unit, signal-connected to the neural network coding unit; an encryption unit, signal-connected to the quantization compression unit; a first communication unit, signal-connected to the encryption unit; and a generative adversarial network (GAN) unit, signal-connected to both the image segmentation unit and the neural network coding unit, the GAN unit including an autoencoder, a decoder, and a discriminator; a second device, including: a second communication unit, signal-connected to the first communication unit; a decryption unit, signal-connected to the second communication unit; a decompression unit, signal-connected to the decryption unit; a neural network decoding unit, signal-connected to the decompression unit; and an image combination unit, signal-connected to the neural network decoding unit; wherein... The database contains a plurality of image data. The image selection unit selects an image data from the database and transmits the selected image data to the image segmentation unit. The image segmentation unit segments the image data to generate a plurality of segmented image data. The image data can be an image, video, video stream, audio data, or compressed image data. Each segmented image data is a fixed-size block, and each block is mapped to a vector. These vectors are considered as a series of image words. The image unit is the segmented image data. The neural network coding unit encodes the segmented image data to generate encoded image data. The quantization compression unit quantizes and compresses the encoded image data to generate compressed image data. The encryption unit encrypts the compressed image data to generate encrypted image data. The decryption unit decrypts the encrypted image data to generate decrypted image data. The decompression unit decompresses the decrypted image data. The neural network decoding unit decodes the decompressed image data to generate decoded image data. The image combining unit combines the decoded image data to generate reconstructed image data. Alternatively, the segmented image data is transmitted to the generative adversarial network unit. The autoencoder extracts and quantizes features using an analytical convolutional network at the encoding end, and reconstructs the image data using a synthetic convolutional network at the decoding end. An entropy model is introduced to predict the probability distribution of latent variables for entropy coding. The discriminator judges the entropy coding and provides feedback to improve the perceptual quality of the decoded image. The autoencoder discards more data, and the decoder's generation capability guesses and restores texture details to achieve a higher compression ratio and generate feature data. This feature data is provided to the neural network encoding unit to obtain a low-dimensional latent variable tensor. The latent variables are equivalent to the transform coefficients in the encoding. Therefore, the neural network encoding unit adaptively learns the latent variables and can effectively represent the content of the image data.

2. The AI-driven image data compression and decompression system as described in claim 1 further includes a generative adversarial network (GAN) unit, wherein the GAN unit is signal-connected to the image segmentation unit and the neural network coding unit.

3. The AI-driven image data compression and decompression system as described in claim 1, wherein the first device or the second device is a computer, tablet computer, handheld communication device, closed-circuit surveillance system, vehicle-mounted device, unmanned vehicle, drone, unmanned vehicle, unmanned boat, robot or laptop computer.

4. An AI-driven method for compressing and decompressing image data, comprising: Step S01, selection: a database contains a plurality of image data, and the image selection unit selects the image data from the database; Step S02, segmentation: the image selection unit transmits the image data to an image segmentation unit, and the image segmentation unit segments the image data to generate a plurality of segmented image data, wherein the image data is an image, video, video stream, audio data, or compressed image data, the segmented image data is a fixed-size block, each block is mapped to a vector, these vectors are regarded as a series of image words, and the image unit is the segmented image data; Step S03, Encoding: The image segmentation unit transmits the segmented image data to the neural network coding unit, which encodes the segmented image data to generate encoded image data; or, the segmented image data is transmitted to the generative adversarial network unit, where the autoencoder at the encoding end uses an analytical convolutional network to extract and quantize features, and at the decoding end uses a synthetic convolutional network to reconstruct the image data, and introduces an entropy model to predict the probability distribution of latent variables for entropy coding. The discriminator of the generative adversarial network unit judges the entropy coding and provides feedback to improve the perceptual quality of the decoded image. The autoencoder discards more data, and the decoder's generation capability guesses and restores texture details to obtain a higher compression ratio to generate feature data. This feature data is provided to the neural network coding unit to obtain a low-dimensional latent variable tensor. The latent variables are equivalent to the transformation coefficients in the encoding, so the neural network coding unit adaptively learns the latent variables and can effectively represent the content of the image data. Step S04, Compression: The neural network coding unit transmits the coded image data to the quantization compression unit, which compresses the coded image data to generate compressed image data. Step S05, Encryption: The quantization compression unit transmits the compressed image data to the encryption unit, which encrypts the compressed image data to generate encrypted image data. Step S06, Transmission: The encryption unit transmits the encrypted image data to the first communication unit, which then transmits it to the second communication unit. Step S07, Decryption: The second communication unit transmits the encrypted image data to the decryption unit, which decrypts the encrypted image data to generate decrypted image data. Step S08, Decompression: The decryption unit transmits the decrypted image data to the decompression unit, which decompresses the decrypted image data to generate decompressed image data. Step S09, Decoding: The decompression unit transmits the decompressed image data to the neural network decoding unit, which decodes the decompressed image data to generate decoded image data.Step S10, Reconstruction: The neural network decoding unit transmits the decoded image data to the image combining unit, which combines the decoded image data to generate reconstructed image data.

5. The AI-driven image data compression and decompression method as described in claim 5 further includes a training method for the neural network coding unit, the training method comprising: Step SA, initial selection: the image selection unit selects any image data from the database; Step SB, initial segmentation: the image selection unit transmits the image data to the image segmentation unit, the image segmentation unit segments the image data to generate a plurality of segmented image data; Step SC, extraction: the image segmentation unit transmits the segmented image data to the generative adversarial network unit, the generative adversarial network unit extracts features from the segmented image data to generate feature data; Step SD, training: the generative adversarial network unit transmits the feature data to the neural network coding unit, the neural network coding unit trains based on the feature data.

Images