Neural Picture Filtering with Data Rearrangement for Distribution Mismatch
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network filters in video processing often exhibit non-ideal filtering effects due to differences between the data distribution of training and actual information, leading to suboptimal performance in video compression and decoding.
Innovation Solution
Implement data rearrangement modes before filtering with a neural network filter to align the data distribution of the to-be-filtered information with that of the training data, followed by inverse rearrangement to improve filtering efficacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a neural network filter is used for picture filtering, then filtering performance is improved, but the data distribution difference between training data and actual data leads to non-ideal filtering effects
Solution Approach 1:
The patent applies preliminary action by performing data rearrangement on the reconstructed picture block before feeding it into the neural network filter. This rearrangement operation transforms the data distribution of the actual picture to match the training data distribution, ensuring the neural network processes input data in the same statistical state as during training, thereby maintaining optimal filtering performance.
Solution Approach 2:
The patent changes the data distribution parameter by applying rearrangement operations (such as row shuffling, column shuffling, or block permutation) to transform the statistical properties of the input data. This parameter transformation aligns the data distribution characteristics of the actual picture with those of the training data, resolving the distribution mismatch issue.
2Adaptability or versatility
If data rearrangement is performed before filtering, then data distribution alignment with training data is improved, but additional processing steps are introduced
Solution Approach 1:
The patent segments the picture into smaller blocks (such as 8x8 or 16x16 pixel blocks) and applies rearrangement operations to these segments independently. This segmentation approach simplifies the overall processing complexity by breaking down the large picture into manageable units that can be processed more efficiently, while still achieving the data distribution alignment objective.
Data Source
AI summary
This application provides a neural network-based picture filtering method performed by an electronic device. The method includes: determining a first data processing mode for to-be-filtered information; performing, when the first data processing mode is a data rearrangement mode, data rearrangement on a reconstructed picture block in the to-be-filtered information according to the first data processing mode, and filtering rearranged to-be-filtered information by using a neural network filter, to obtain a filtered picture block of the rearranged to-be-filtered information; and performing data inverse rearrangement on the filtered picture block of the rearranged to-be-filtered information according to the first data processing mode, to obtain a filtered reconstructed picture block. In this application, data rearrangement is performed on the to-be-filtered information, so that a data distribution characteristic of the rearranged to-be-filtered information is close to a data distribution characteristic of training data, improving a filtering effect of a neural network filter.