Image Compression System Using Weighted Feature Map Downsampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image compression techniques using deep learning face challenges in reducing decoding time and achieving accurate probability distribution prediction, particularly due to bottlenecks in potential variables and high demands for code streams.
Innovation Solution
An image compression system that includes an image coding apparatus for downsampling input images, a probability model generating apparatus using a separate entropy model to predict probability distributions, and an image decoding apparatus for upsampling, which reduces decoding time by weighting and down-dimension processing feature maps, and achieves accurate probability distribution prediction using a context and entropy model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning techniques are used for image compression, then compression capability is improved, but decoding time increases due to bottlenecks in potential variables
Solution Approach 1:
The patent divides the probability model generation into two independent segments: context model processing and entropy model processing. The context model generates content-based predictions, while the entropy model combines these with auxiliary information to produce the final probability model. This segmentation allows parallel processing and reduces the sequential bottleneck in decoding operations.
Solution Approach 2:
The patent introduces a hyper decoder as an intermediary component that generates auxiliary information independently of the main decoding path. This auxiliary information serves as a mediator that enhances probability prediction accuracy without increasing the complexity of the main decoding pipeline, thus reducing decoding time while maintaining compression capability.
2Productivity
If accurate probability distribution prediction is achieved, then compression efficiency is improved, but code stream demands increase
Solution Approach 1:
The patent applies partial action by using a hyper decoder to generate only the necessary auxiliary information for probability prediction, rather than processing the entire image data through the full deep learning pipeline. This selective processing achieves accurate probability distribution prediction while minimizing the amount of data that needs to be transmitted in the code stream.
Solution Approach 2:
The patent changes the parameter representation by separating the probability model into mean and variance portions, each processed through different models (context model and entropy model). This parameter transformation allows more efficient encoding by matching the statistical properties of different image regions with appropriate prediction strategies, reducing overall code stream requirements.
3Manufacturing precision
If feature maps of N channels are processed to maintain detail, then image quality is preserved, but processing complexity increases
Solution Approach 1:
The patent dynamically adjusts the processing of feature maps by applying different operations to different channels based on their importance. The weighting unit assigns different weights to different channels, and the down-dimension processing selectively reduces channels while preserving those with higher weights. This dynamic approach maintains image quality for important features while reducing processing complexity for less critical channels.
Solution Approach 2:
The patent applies local quality by treating different channels of feature maps differently rather than uniformly. The weighting mechanism identifies which channels contain more important information and preserves them with higher fidelity, while allowing greater compression for less important channels. This localized differentiation maintains overall image quality while reducing total processing complexity.
Data Source
AI summary
Embodiments of this disclosure provide an image coding method and apparatus and an image compression system. The image coding apparatus includes a memory and a processor. The processor is configured to perform feature extraction on an input image to obtain feature maps of N channels; assign a weight to a feature map of each channel among the N channels; perform down-dimension processing on weighted feature maps processed in association with the N channels, to obtain feature maps of M channels and output the feature maps of M channels, M being smaller than N. Hence, by multiplying different feature maps by a weight to obtain corresponding importance and then performing down-dimension processing on the feature maps processed according to the weighting, time for decoding may be reduced.