Neural Deep Intra Prediction for Image Blocks With Masked Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression systems like HEVC and VVC face challenges in distinguishing missing pixels from available pixels in the context during intra prediction of image blocks, leading to inefficiencies and increased memory usage.
Innovation Solution
Utilizing a neural network-based approach with a mask value or binary vector to distinguish missing pixels from available pixels in the context, followed by pre- and post-processing to clarify ambiguities and enhance prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network-based intra prediction approach is used, then prediction accuracy is improved, but the ability to distinguish missing pixels from available pixels deteriorates due to ambiguities in the context
Solution Approach 1:
The patent introduces mask values as an intermediary element that mediates between the context pixels and the neural network. These mask values are inserted into the context at positions corresponding to missing pixels, serving as a bridge that communicates pixel availability information to the neural network without requiring separate signaling mechanisms.
Solution Approach 2:
The patent changes the parameter representation of pixel availability by transforming the binary state (present/absent) into a continuous mask value that can be directly integrated into the neural network's input tensor. This allows the network to process availability information as part of the normal feature extraction process rather than as a separate control parameter.
2Measurement precision
If the context includes all surrounding pixels, then prediction quality is improved, but memory usage increases due to storing and processing incomplete pixel data
Solution Approach 1:
The patent extracts only the necessary pixel information from the complete context by identifying and removing missing pixels. Instead of processing the entire surrounding area including gaps, the method extracts a compact representation that includes only available pixels plus their corresponding mask values, reducing the data volume that needs to be stored and processed.
Solution Approach 2:
The patent segments the context into individual pixel positions, each treated as a separate element that can be independently evaluated for availability. This segmentation allows the mask values to be precisely placed at missing pixel locations, enabling the neural network to process a structured representation that separates present and absent data points.
Data Source
AI summary
There is provided intra predicting of an image block using at least one neural network from a context comprising pixel surrounding said image block. Said intra predicting also uses an information to distinguish the missing pixels from the available pixels in the context. There are also provided encoding/decoding methods and apparatus, a device, a non-transitory computer readable medium and a computer program product.