Feature Map Decoding with Distribution Expansion for Vision Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The distribution of reconstructed images or feature maps differs from the original images or feature maps when encoded/decoded by an artificial neural network, leading to a decline in machine vision performance.
Innovation Solution
An image feature map encoding/decoding method based on latent representation distribution expansion, using distribution expansion parameters to correct the distribution of reconstructed feature maps, which includes obtaining a reconstructed feature map by decoding a feature map latent representation and performing distribution expansion based on these parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature map latent representation is encoded and decoded by an artificial neural network, then compression and transmission are achieved, but the distribution of reconstructed feature maps differs from original feature maps causing performance degradation
Solution Approach 1:
The patent applies parameter changes by introducing distribution expansion parameters (scale parameter s and shift parameter t) that modify the statistical distribution of reconstructed latent representations. These parameters are learned and applied to adjust the mean and variance of the reconstructed features, transforming them to match the original feature distribution and thereby resolving the performance degradation caused by distribution mismatch during compression and decoding
Solution Approach 2:
The patent introduces distribution expansion parameters as intermediary elements between the decoder output and the final reconstructed feature map. These parameters act as mediators that bridge the distribution gap between compressed and original features, allowing the system to maintain both compression efficiency and reconstruction accuracy by adjusting the intermediate latent representation distribution
2Reliability
If distribution expansion is applied to correct reconstructed feature map distribution, then machine vision performance is improved, but additional parameters and computational steps are required
Solution Approach 1:
The patent applies preliminary action by pre-learning the distribution expansion parameters (scale s and shift t) during the training phase. These parameters are computed in advance based on the statistical properties of the training data and are stored for use during decoding, eliminating the need for complex real-time distribution adjustment operations and reducing online computational complexity
Solution Approach 2:
The patent simplifies the complexity issue by using parameter changes - specifically, by representing the distribution adjustment through simple scale and shift parameters that can be applied through basic arithmetic operations rather than complex transformations. This maintains performance improvement while keeping the additional computational overhead minimal
Data Source
AI summary
An image feature map encoding/decoding method, device and recording medium based on the latent representation distribution expansion of the present disclosure may include obtaining a reconstructed feature map latent representation by decoding a feature map latent representation obtained by encoding a feature map from a bitstream, and obtaining a reconstructed feature map by decoding the reconstructed feature map latent representation, wherein the reconstructed feature map may be obtained by performing distribution expansion based on a distribution expansion parameter obtained from the bitstream.