Neural Network Loop Filtering with Quantization Step Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression technologies face challenges in achieving effective filtering for reconstructed pictures of varying quality levels, particularly in hybrid video encoding and decoding systems, where neural networks are used for loop filtering but fail to provide a good filtering effect across different quality levels.
Innovation Solution
A loop filtering method that utilizes a neural network to process input pixel matrices, including both luminance and quantization step values, to improve filtering performance by introducing quantization step values for each pixel, allowing for better filtering processing and enhancing the filtering effect across various quality levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a neural network is used for loop filtering in hybrid video encoding and decoding, then filtering processing can be performed on reconstructed picture information, but the filtering effect is poor for input pictures or picture blocks of different quality levels
Solution Approach 1:
The patent introduces quantization step values as additional input parameters to the neural network, alongside pixel values. This changes the parameter set fed into the network, enabling it to adapt its filtering behavior based on the quality characteristics (quantization steps) of different input pictures or picture blocks, thereby improving filtering effect across varying quality levels.
Solution Approach 2:
The patent performs preliminary extraction and preparation of quantization step values from the encoded video data before feeding them into the neural network. This preliminary action allows the network to have advance knowledge of the quality characteristics of the input data, enabling it to adjust its filtering strategy accordingly and achieve better filtering effects for different quality levels.
2Quantity of substance
If video data is compressed to reduce data amount for transmission or storage, then bandwidth usage and storage requirements decrease, but picture quality may be sacrificed
Solution Approach 1:
The patent uses a neural network that takes both pixel values and quantization step values as input and produces filtered reconstructed picture blocks as output. This creates a feedback mechanism where the filtering process is adaptive and informed by the actual quality characteristics of the compressed data, allowing the system to improve picture quality without requiring additional compression data, thus maintaining quality while keeping data amount reduced.
Data Source
AI summary
A loop filtering method and apparatus are provided. The method includes: obtaining a first pixel matrix, where a value of a pixel at a corresponding location in the first pixel matrix corresponds to a luminance value of a pixel at a corresponding location in a first picture block; obtaining a second pixel matrix, where a pixel at a corresponding location in the second pixel matrix corresponds to a quantization step value corresponding to the luminance value of the pixel at the corresponding location in the first picture block; performing filtering processing on an input pixel matrix by using a filtering network to obtain an output pixel matrix, where the filtering network is a neural network that has a filtering function and is obtained through training, the output pixel matrix includes a third pixel matrix. Thus, filtering effect can be improved for reconstructed pictures of various quality levels.