Decoder-Side Intra Mode Merge Using Gradient Histograms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current implementations of Decoder Side Intra Mode Derivation (DIMD) and DIMD merge in video coding are limited by considering only spatial information and are applied only when neighboring blocks are encoded with DIMD or DIMD Merge, which restricts their performance.
Innovation Solution
A method and device that obtain histograms of gradients from all available neighboring blocks regardless of their encoding mode, merge these gradients to determine the highest cumulative magnitudes, and compute a weighted sum of predictors using planar modes to improve intra prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If DIMD merge considers only spatial information from neighboring blocks encoded with DIMD or DIMD Merge, then the implementation complexity is reduced, but the intra prediction performance is limited
Solution Approach 1:
The patent extends the applicability of DIMD merge by allowing it to function with multiple types of neighboring blocks (spatially neighboring blocks regardless of encoding mode, and collocated blocks from reference pictures). This multi-functionality enables the system to utilize both spatial and temporal information, improving prediction accuracy without significantly increasing implementation complexity since the same HoG computation and merging mechanisms are used.
Solution Approach 2:
The patent introduces a temporal dimension to the traditional spatial-only DIMD merge by incorporating collocated blocks from reference pictures. This adds a new dimension (time/reference picture dimension) to the information gathering process, allowing the system to exploit temporal correlations in addition to spatial correlations, thereby improving prediction performance.
2Productivity
If DIMD merge is applied only when neighboring blocks are encoded with DIMD or DIMD Merge, then the computational overhead is reduced, but the versatility of the method is limited
Solution Approach 1:
The patent makes the DIMD merge method more universal by enabling it to work with any neighboring block regardless of its encoding mode (intra or inter), and by allowing the use of collocated blocks from reference pictures. This universality is achieved through the same computational mechanism (HoG computation and merging), maintaining encoding efficiency while greatly expanding method applicability.
Solution Approach 2:
The patent introduces dynamic adaptability by allowing the system to selectively use different types of neighboring blocks (spatial or temporal) and different encoding modes based on availability and suitability. The decoder can dynamically choose the most appropriate source of HoG information for each current block, making the method more adaptable to various coding scenarios.
3Measurement precision
If histograms of gradients are computed for all neighboring blocks regardless of encoding mode, then the prediction performance is improved, but the computational complexity increases
Solution Approach 1:
The patent merges HoG information from multiple sources (spatially neighboring blocks and collocated blocks) using the same computational mechanism. By combining temporal and spatial information through the same HoG computation and merging process, the system achieves improved prediction accuracy without requiring fundamentally different computational approaches, thus limiting the increase in computational complexity.
Solution Approach 2:
The patent reuses the same HoG computation and merging mechanism for both spatial and temporal dimensions. Instead of developing separate complex algorithms for temporal prediction, the system copies and applies the proven spatial DIMD merge mechanism to temporal information from collocated blocks, achieving performance improvement with minimal additional computational burden.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for encoding a current block comprising: obtaining (80 the current block; obtaining (1101) an histogram of gradients from at least each available spatially neighboring block of the current block regardless of the mode used to encode this spatially neighboring block, each obtaining of an histogram of gradients for a spatially neighboring block comprising computing the histogram of gradients responsive to no histogram of gradients had been already computed for the neighboring block; determining (1102) a merged histogram of gradients from each obtained histogram of gradients; selecting (83) at least one direction with highest cumulative magnitude in the first merged histogram of gradients to obtain at least one Decoder Side Intra Mode Derivation (DIMD) modes; determining (84) a weight for each DIMD mode obtained from the selecting of at least one direction, each weight depending on relative magnitudes in the first merged histogram of gradients; and, computing (85) a weighted sum of predictors obtained from the at least one DIMD mode and for a planar mode using each determined weights to obtain a final predictor for the current block.