Feature Encoding and Decoding for AI Image Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression technologies are not optimized for artificial intelligence services, leading to inefficiencies in processing and reconstructing feature maps due to feature point dimension reduction, which affects encoding and decoding efficiency and accuracy.
Innovation Solution
A feature encoding/decoding method and apparatus that applies a filter to minimize the difference between the original and reconstructed feature maps, transmitting filtering information to compensate for errors caused by dimension reduction, and storing or transmitting bitstreams generated by the encoding method or apparatus.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature point dimension reduction is applied to compress feature maps, then compression efficiency is improved, but reconstruction accuracy deteriorates due to errors introduced by dimension reduction
Solution Approach 1:
The patent applies filtering operations before dimension reduction and after reconstruction to preemptively minimize errors. The encoder applies a filter to the original feature map before compression, and the decoder applies filtering to the reconstructed feature map, preparing the data in advance to reduce reconstruction errors rather than correcting them after they occur.
Solution Approach 2:
The patent transmits filter information from the encoder to the decoder through the bitstream, creating a feedback mechanism. The decoder uses this transmitted filter information to apply the same filtering operations, ensuring that both encoder and decoder use consistent filtering parameters to minimize reconstruction errors and maintain accuracy.
2Measurement precision
If filtering operations are applied to minimize differences between original and reconstructed feature maps, then reconstruction accuracy is improved, but device complexity increases
Solution Approach 1:
The patent modifies the encoding/decoding process by introducing filter parameters that can be adjusted to optimize the balance between complexity and accuracy. The filter information is transmitted in the bitstream, allowing the system to adapt filtering strength and type based on specific requirements, thereby controlling complexity while maintaining reconstruction accuracy.
3Productivity
If existing image compression technology is used for AI services, then encoding efficiency is reduced, but compatibility with standard formats is maintained
Solution Approach 1:
The patent segments the feature map processing into distinct stages: pre-filtering, dimension reduction, and post-reconstruction filtering. This segmentation allows each stage to be optimized independently for AI workloads, improving overall encoding efficiency while maintaining compatibility with standard compression formats through the use of conventional dimension reduction techniques in the middle stage.
Data Source
AI summary
Provided are a feature encoding/decoding method and apparatus, and a computer-readable recording medium generated by the feature encoding method are provided. The feature decoding method according to the present disclosure may comprise the steps of: acquiring, from a bitstream, first feature information and filtering information regarding the first feature information; inversely transforming the first feature information; and deriving, on the basis of the filtering information, second feature information for the inversely transformed first feature information.