Feature Data Compression Using Pixel Unshuffle and Non-Linear Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding standards face challenges in efficiently compressing feature data for object tracking, particularly in reducing noise in reconstructed feature data and optimizing encoding for distributed object recognition tasks.
Innovation Solution
The proposed solution involves a method of compressing feature data by performing spatial down-sampling using a pixel unshuffle operation and channel reduction through a non-linear two-dimensional convolution with an activation function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature data is compressed using conventional video coding standards, then data transmission efficiency is improved, but noise in reconstructed feature data increases and object tracking accuracy deteriorates
Solution Approach 1:
The patent segments the feature data compression process into distinct stages: spatial down-sampling using pixel unshuffle operation followed by channel reduction through non-linear two-dimensional convolution. This segmentation allows each stage to be optimized independently, maintaining object tracking accuracy while achieving compression.
Solution Approach 2:
The patent introduces an intermediary processing stage that uses non-linear two-dimensional convolution with activation functions as a mediator between compression and reconstruction. This intermediary operation preserves essential feature information while reducing noise, enabling accurate object tracking from compressed data.
2Quantity of substance
If spatial down-sampling is performed using conventional methods, then data volume is reduced, but noise in reconstructed feature data increases
Solution Approach 1:
The patent replaces conventional linear down-sampling methods with a non-linear approach using two-dimensional convolution and activation functions. This substitution transforms the mechanical down-sampling process into an intelligent feature extraction process that reduces noise while preserving essential information.
Solution Approach 2:
The patent changes the parameters of the down-sampling operation by applying non-linear activation functions and convolution kernels. These parameter changes enable the system to adaptively preserve important feature information during compression, maintaining quality while reducing data volume.
3Productivity
If channel reduction is applied to compressed feature data, then encoding efficiency is improved, but object recognition accuracy may deteriorate
Solution Approach 1:
The patent performs preliminary channel reduction on the compressed feature data before final encoding. This preliminary action prepares the data in an optimized format that maintains essential object recognition information while reducing complexity, enabling efficient subsequent encoding without accuracy loss.
Solution Approach 2:
The patent applies two-dimensional convolution operations that operate across spatial and channel dimensions simultaneously. This dimensional approach allows efficient compression by exploiting correlations in multiple dimensions while preserving the three-dimensional structure of feature data necessary for accurate object recognition.
Data Source
AI summary
A method of compressing feature data includes: receiving feature data; performing spatial down sampling on the received feature data by applying a pixel unshuffle operation; and performing channel reduction on the spatially down sampled feature data by applying a non-linear two dimensional convolution with an activation.