IMAGE DECODING METHOD, IMAGE ENCODING METHOD, IMAGE ENCODING DEVICE, AND IMAGE DECODING DEVICE USING AI

Smart overview of the Invention

The described technology pertains to methods and devices for image encoding and decoding using artificial intelligence (AI), specifically neural networks. The approach involves processing feature data from bitstreams to reconstruct images. This includes handling optical flow and residual image data through neural network-based decoders, which helps in predicting and reconstructing current images from previously reconstructed ones.

Background

Traditional codecs like H.264 and HEVC encode images by dividing them into blocks and using prediction techniques to compress data. Intra prediction removes spatial redundancy, while inter prediction focuses on temporal redundancy. Motion estimation coding is a common inter prediction method, using reference images to predict current image blocks. The need for more efficient AI-based encoding and decoding methods has emerged to enhance these processes.

Decoding Process

The decoding method involves extracting feature data of optical flow and residual images from bitstreams. A neural network-based decoder processes this data to obtain current optical flow and multiple remembering gate values for different image resolutions. These elements are used to predict the current image from previously reconstructed images, generating prediction tensors that aid in reconstructing the current image.

Encoding Process

The encoding method mirrors the decoding process by first obtaining feature data of the current optical flow through a neural network-based encoder. This data, along with remembering gate values, facilitates the prediction of current images. The process generates prediction tensors, which are used alongside the current image in a second encoder to derive feature data of residual images. A bitstream is then created, encapsulating the processed feature data.

Innovation and Applications

This AI-driven method enhances image encoding and decoding by leveraging neural networks to manage data more effectively. It introduces remembering and forgetting gate values to maintain or remove information, optimizing the prediction and reconstruction of images across varying resolutions. This technology offers potential improvements in video compression standards, potentially impacting streaming and storage efficiency.