Feature Map Encoding for Machine Vision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression technologies focus on human vision, which is inadequate for machine vision applications, where the goal is to maximize machine task performance rather than image quality.
Innovation Solution
A method and device for encoding and decoding feature maps using a learned neural network parameter, which extracts and restores multi-layer feature maps from latent representations, allowing for efficient compression and decompression suitable for machine vision tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional image compression technology is used, then image quality for human vision is improved, but machine task performance deteriorates
Solution Approach 1:
The patent applies local quality by treating different types of images (natural images for human vision vs. feature maps for machine vision) with different compression strategies. The system selectively applies appropriate compression methods based on the intended use, ensuring optimal quality for human vision tasks while preserving machine task performance through feature-map-specific compression techniques
Solution Approach 2:
The patent changes the fundamental parameters of the compression approach by transitioning from pixel-value-based compression to feature-based compression. It modifies compression parameters to prioritize preservation of task-relevant features rather than visual fidelity, using learned parameters from neural networks to guide the compression process for machine vision applications
2Quantity of substance
If feature map compression is implemented, then data transmission and storage are reduced, but compression ratio and quality optimization become more complex
Solution Approach 1:
The patent applies preliminary action by pre-processing feature maps through neural network extraction before compression. The system performs feature extraction and latent representation learning in advance, transforming raw feature maps into compressed latent forms that are then easier to compress efficiently, reducing the overall complexity of the compression pipeline
Solution Approach 2:
The patent introduces an intermediary latent representation layer between the original feature map and the compressed form. This latent space acts as a mediator that captures essential features while reducing dimensionality, simplifying the compression process by working with intermediate representations rather than raw high-dimensional feature maps
Data Source
AI summary
A device for decoding a feature map according to the present disclosure comprises an image decoding unit to decode an image from a bitstream; an inverse format conversion unit to restore a feature map latent representation by converting a formation of a decoded image; and a feature map restoration unit to restore a multi-layer feature map from the feature map latent representation. Here, the feature map restoration unit restores the multi-layer feature map based on a learned neural network parameter.