Multi-Resolution Feature Map Encoding with Residual Compensation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image encoding/decoding technologies face challenges in effectively compensating for compression damage in high-resolution feature maps, leading to degradation in image quality during transmission and storage.
Innovation Solution
The method involves extracting feature maps from an input image, performing packing and encoding on these maps, and using residual values to compensate for compression damage, specifically through upsampling techniques like double cubic, bilinear, or deep learning-based interpolation methods, to generate a bitstream that can be stored and transmitted.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If compression encoding is performed on feature maps for transmission and storage, then transmission efficiency and storage capacity are improved, but compression damage occurs leading to degradation in image quality
Solution Approach 1:
The patent divides the feature map into multiple resolutions (high-resolution and low-resolution feature maps) and processes them differently. The low-resolution feature map is used for generating residual values that compensate for compression damage, while the high-resolution feature map maintains image quality. This segmentation allows efficient compression while preserving critical quality information.
Solution Approach 2:
The patent applies different quality levels to different parts of the feature map by generating residual values from low-resolution versions and using these to compensate for compression damage in the high-resolution feature map. This local quality approach ensures that compression damage is compensated where most needed while maintaining overall efficiency.
2Manufacturing precision
If high-resolution feature maps are transmitted without compression compensation, then image quality is maintained, but transmission bandwidth and storage requirements increase significantly
Solution Approach 1:
The patent extracts residual values from the low-resolution feature map that represent compression damage, and separately transmits these residual values along with the compressed high-resolution feature map. This extraction allows the receiver to compensate for compression damage without transmitting the full uncompressed high-resolution data, thereby reducing bandwidth requirements while maintaining image quality.
3Manufacturing precision
If residual values are calculated and used for compensation, then compression damage is reduced, but computational complexity and processing time increase
Solution Approach 1:
The patent dynamically adjusts the compensation process by using upsampling operations to scale the residual values from low-resolution to high-resolution. This dynamic approach allows the system to adapt the compensation level based on the specific compression applied, balancing computational complexity with compensation effectiveness.
Solution Approach 2:
The patent changes the resolution parameter of the residual values through upsampling operations (such as bilinear or bicubic interpolation) to match the high-resolution feature map. This parameter transformation enables efficient compensation without requiring complex processing, as the upsampling operations are computationally lightweight compared to full-resolution processing.
Data Source
AI summary
Disclosed herein are a method, an apparatus, and a storage medium for encoding/decoding a multi-resolution feature map. In an aspect, there is provided an encoding method, including extracting feature maps from an input image, performing packing on the extracted feature maps, and performing encoding on the packed feature maps, wherein the feature maps include multiple feature maps having different resolutions. In another aspect, there is provided a decoding method, including generating packed feature maps by decoding an input bitstream, and generating a multi-resolution feature map by performing unpacking on the packed feature maps, wherein the multi-resolution feature map includes multiple feature maps having different resolutions.