Neural Image Compression with Adaptive Encoder-Decoder Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression and reconstruction technologies using neural network models face challenges in achieving efficient data compression while adhering to footprint constraints, particularly in systems-on-chip (SoC) environments, where power consumption and bandwidth are critical.
Innovation Solution
Employing a system comprising multiple encoders and decoders with different characteristics, trained on varying loss functions, to select optimal encoder-decoder pairs for each image block region, combined with entropy encoding to generate and decode vector values, thereby optimizing compression performance and reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple encoders and decoders are used to improve compression performance, then data compression ratio and quality are improved, but device complexity and footprint increase
Solution Approach 1:
The patent divides the image into multiple regions and applies different encoder-decoder pairs to different regions based on their characteristics. This segmentation allows the system to use multiple specialized models only where needed, improving compression performance for specific region types while limiting the overall footprint by not applying all models uniformly across the entire image.
Solution Approach 2:
The system dynamically selects appropriate encoder-decoder pairs for different image regions based on their characteristics. Rather than using a fixed single model or all models simultaneously, the selection is adaptive and context-dependent, allowing the system to optimize between compression ratio and footprint by choosing only the necessary models for each specific region.
2Manufacturing precision
If multiple encoders and decoders with different loss functions are used, then reconstruction quality is improved, but power consumption increases
Solution Approach 1:
Different encoder-decoder pairs with different loss functions are applied to different image regions based on their specific characteristics. This local quality approach ensures that high reconstruction quality is achieved only where necessary, while reducing power consumption by using simpler or different models in regions where they are not needed.
Solution Approach 2:
The system dynamically selects which encoder-decoder pair to use for each region based on the input data characteristics. This dynamic selection allows the system to optimize power consumption by choosing models with appropriate complexity for each specific region, rather than uniformly applying all models or a single high-complexity model throughout.
3Productivity
If encoder selection is performed based on input data characteristics, then compression efficiency is improved, but processing time increases
Solution Approach 1:
The system performs preliminary analysis of input data characteristics to determine which encoder-decoder pair is most suitable before actual compression. This preliminary action allows for optimized compression efficiency by selecting the best model in advance, while the processing time impact is managed by making the selection process efficient and region-based rather than applying complex selection to the entire image.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A data processing method may comprise receiving input data, selecting an encoder to encode the input data among a plurality of encoders based on the input data, obtaining a vector value by encoding the input data using the selected encoder, and generating the encoded data comprising the vector value, and identification data that identifies a decoder to decode the vector value among a plurality of decoders.