Feature Frame Compression for Multi-Channel Feature Map Transmission
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of efficiently compressing large-capacity multi-channel feature maps generated by deep learning networks for use on personal/client devices with limited memory and computation resources, while maintaining task performance, is not adequately addressed by existing technologies.
Innovation Solution
A method and apparatus for encoding and decoding feature maps using a feature frame configuration and compression technique, involving normalization, sorting, and conversion to facilitate efficient transmission and reconstruction, utilizing video compression codecs and deep learning-based image compression methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If feature maps are transmitted to personal/client devices for deep learning network execution, then task performance can be achieved on devices with limited resources, but the data transmission size becomes excessively large
Solution Approach 1:
The patent segments the deep learning network into two parts: a server-side network for feature extraction and a client-side network for task execution. Only the necessary feature maps are transmitted to the client device, rather than the entire network, enabling task performance on devices with limited resources while reducing data transmission size.
Solution Approach 2:
The patent extracts and transmits only the essential feature maps from the server-side deep learning network to the client device. This extraction approach allows the client device to perform tasks using minimal data, achieving task performance without requiring large data transmissions or storing the complete network model.
2Reliability
If multiple feature maps are generated for different channels, then comprehensive feature representation is achieved, but the compression and transmission burden increases
Solution Approach 1:
The patent merges multiple feature maps into a unified feature frame structure that can be efficiently compressed and transmitted. By organizing feature maps into this consolidated format, the system maintains comprehensive feature representation while reducing the overall compression and transmission burden compared to handling each feature map separately.
Solution Approach 2:
The patent applies parameter changes through normalization and quantization processes to reduce the precision requirements of feature map data. This allows the system to maintain sufficient feature representation quality while significantly reducing the data size for compression and transmission, thereby lowering the overall system complexity.
3Quantity of substance
If feature maps are compressed to reduce transmission size, then data efficiency improves, but compression performance and reconstruction quality deteriorate
Solution Approach 1:
The patent employs parameter changes through normalization and quantization techniques to compress feature maps while preserving essential information. By carefully adjusting these parameters, the system achieves effective compression that reduces transmission size while maintaining sufficient reconstruction quality for task performance.
Solution Approach 2:
The patent incorporates feedback mechanisms that allow the system to evaluate compression performance and adjust parameters accordingly. This feedback loop ensures that compression operations maintain adequate reconstruction quality while achieving the desired reduction in data transmission size, balancing efficiency and performance.
Data Source
AI summary
Disclosed herein are a method, an apparatus and a storage medium for processing a feature map. An encoding method for a feature map includes configuring a feature frame for feature maps, and generating encoded information by performing encoding on the feature frame. A decoding method for a feature map includes reconstructing a feature frame by performing decoding on encoded information, and reconstructing feature maps using the feature frame. A feature frame is configured using feature maps, and compression using a video compression codec or a deep learning-based image compression method is applied to the feature frame.