Neural Network Feature Map Encoding via Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in efficiently encoding feature maps extracted from artificial neural networks while minimizing data amount and maintaining task performance, particularly when the feature map resolution needs to be reduced without degrading small object detection rates.
Innovation Solution
The proposed solution involves converting the resolution of feature maps using a super-resolution technique, applying compression and reconstruction, and learning from the results to improve encoding performance by aligning channel sizes with encoding block sizes and using metadata for feature map reconstruction modes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the resolution of the feature map is decreased by selecting an encoding feature map having lower resolution than the original feature map, then the data amount of the feature map is reduced, but the detection rate of small objects may decrease
Solution Approach 1:
The patent applies super-resolution processing to the encoding feature map before compression encoding. This preliminary enhancement of resolution allows the system to work with lower-resolution feature maps (reducing data amount) while still achieving high detection accuracy for small objects after the super-resolution restoration.
Solution Approach 2:
The patent changes the resolution parameter of the encoding feature map to be lower than the original feature map resolution. By combining this parameter change with super-resolution processing, the system reduces the data amount requiring compression while maintaining or improving the detection performance through resolution restoration.
2Productivity
If the feature map is compressed to reduce data amount, then the transmission efficiency is improved, but compression artifacts may degrade the quality of the feature map
Solution Approach 1:
The patent applies super-resolution processing to enhance the encoding feature map before compression encoding. This preliminary enhancement creates a more robust representation that better withstands the subsequent compression process, reducing the degradation of feature map quality while maintaining transmission efficiency.
Solution Approach 2:
The patent uses the original high-resolution feature map as a reference during the super-resolution processing of the encoding feature map. This feedback mechanism ensures that the compressed feature map maintains fidelity to the original features, preserving quality while enabling efficient compression.
Data Source
AI summary
Disclosed herein is an encoding method. The encoding method includes extracting a feature map from an input image, determining an encoding feature map based on the extracted feature map, generating a converted feature map by performing conversion on the encoding feature map, and performing encoding on the converted feature map.