Encoding Method and Apparatus, Decoding Method and Apparatus, Device, Storage Medium, and Computer Program Product

Smart overview of the Invention

This disclosure pertains to encoding and decoding technologies, focusing on methods and apparatuses for encoding and decoding, along with devices, storage mediums, and computer program products. It addresses challenges in image compression, particularly in scenarios involving varying bit rates, by improving the stability and performance of probability distribution estimation networks used in these processes.

Background

Image compression technology is crucial for the efficient transmission and storage of image information, especially as data volumes grow. Traditional methods involve extracting and quantizing image features, then encoding these features using entropy coding. However, quantization steps need to adapt to different bit rates, causing significant variations in the numerical ranges of image features. This variability complicates the training of probability distribution estimation networks, affecting the overall performance of encoding and decoding processes.

Summary of Invention

The invention improves encoding and decoding by simplifying the training of probability distribution estimation networks, even in multi-bit-rate scenarios. It achieves this by estimating the probability distribution of unquantized image features based on hyperprior features. This approach stabilizes the numerical range of unquantized features, making network training more effective and enhancing encoding and decoding performance.

Encoding Method

The encoding method involves determining image features and hyperprior features, encoding these into a bitstream, and estimating probability distribution parameters. By quantizing these parameters, the method encodes image features efficiently. This process ensures that even with varying bit rates, the unquantized image features remain stable, facilitating more reliable network training and improved performance.

Decoding Method

In decoding, the bitstream is parsed to obtain hyperprior features, from which probability distribution parameters are derived. These parameters are quantized to reconstruct the image accurately. The method maintains the stability of unquantized image features, ensuring that the probability distribution estimation network performs effectively, enhancing the overall decoding quality.