Feature Map Compression with Selective Transmission and Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of efficiently transmitting feature maps from a feature pyramid network while minimizing data loss and maintaining performance in machine tasks at the reception edge is addressed.
Innovation Solution
A deep learning-based method for selectively transmitting feature maps and predicting untransmitted maps using a neural network, involving decoding, inverse packing, and reconstructing feature maps with different resolutions, utilizing a convolutional neural network and generalized divisive normalization layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If compression rate is improved in encoder, then data transmission volume is reduced, but compression damage increases and performance degrades
Solution Approach 1:
The patent extracts and transmits only the essential feature maps from the feature pyramid network, selectively choosing which feature maps to transmit based on their importance. This allows reducing data transmission volume while maintaining the most critical information needed for machine tasks at the reception edge.
Solution Approach 2:
The patent uses a neural network to generate compressed representations (copies) of the transmitted feature maps. The reception edge creates approximate copies of the original feature maps through neural network processing, which reduces compression damage while maintaining task performance with minimal data transmission.
2Adaptability or versatility
If multi-resolution feature maps are transmitted, then object detection hardiness is maintained, but data transmission volume increases
Solution Approach 1:
The patent segments the feature pyramid network outputs into multiple resolution levels and selectively transmits only certain segments (feature maps) rather than all of them. This segmentation allows maintaining adaptability for different object sizes while reducing the total data transmission volume by transmitting only the most important resolution levels.
3Productivity
If feature map compression is applied, then transmission efficiency is improved, but compression damage occurs
Solution Approach 1:
The patent employs neural networks to generate compressed copies of feature maps at the reception edge. These neural-generated copies preserve essential information while allowing aggressive compression during transmission, thereby improving transmission efficiency while minimizing information loss through intelligent reconstruction.
Solution Approach 2:
The patent changes the representation parameters of feature maps by transforming them into a compressed latent space using neural networks. This parameter transformation allows efficient transmission with reduced data volume while the neural network reconstruction process recovers the essential information, minimizing compression damage.
Data Source
AI summary
The present disclosure may be a deep learning-based method for improving feature map compression efficiency, wherein feature maps can be selectively transmitted from an encoder, and untransmitted feature maps can be predicted and generated after restoring the image quality of the transmitted feature maps using a deep neural network.