Deep Learning Image Compression with Downsampling Back Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression methods face challenges such as high computational effort in designing prediction modes, noticeable blocking effects at high compression rates, and issues with image blurring, noise, and compression artifacts, particularly in deep learning-based methods.
Innovation Solution
The proposed method employs a deep learning-based image compression technique using a coding network with downsampling back projection modules. These modules perform downsampling, reconstruction, and optimization of feature maps to obtain a hidden variable, which is then used to generate a bitstream file through entropy coding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep learning-based methods are used for image compression, then compression quality is improved, but image blurring, noise, and compression artifacts occur
Solution Approach 1:
The patent implements a feedback mechanism where the reconstruction image is fed back into the coding network to generate a difference image, which is then used to optimize the hidden variable. This feedback loop allows the system to iteratively refine the compression result, reducing blurring, noise, and compression artifacts while maintaining high compression quality.
Solution Approach 2:
The patent changes the parameter representation by introducing an optimized hidden variable that incorporates difference image information. Instead of directly compressing the original image, the system compresses the difference between the original and reconstruction images, allowing for more precise control over compression parameters and reducing harmful artifacts.
2Productivity
If compression rate is increased, then bandwidth efficiency is improved, but blocking effect becomes more obvious
Solution Approach 1:
The patent segments the image compression process into multiple stages: initial compression to achieve high bandwidth efficiency, reconstruction to restore image quality, and difference image generation to capture and compress only the essential information. This segmentation allows the system to achieve high compression rates without obvious blocking effects.
Solution Approach 2:
The patent introduces a reconstruction image as an intermediary element between the compressed hidden variable and the final output. This intermediary allows the system to maintain high bandwidth efficiency while the reconstruction process mitigates blocking effects by providing a reference for quality restoration.
3Productivity
If existing filtering or block prediction methods are used, then compression is achieved, but large amount of time and effort is required for manual design of prediction modes
Solution Approach 1:
The patent replaces manual mechanical design of prediction modes with an automated deep learning-based system. The coding network automatically learns optimal compression representations through training, eliminating the need for manual design of prediction modes and significantly reducing the time and effort required while maintaining compression efficiency.
Solution Approach 2:
The patent changes the approach from manual parameter design to automated parameter learning. Instead of manually designing prediction modes, the system uses neural networks to automatically learn and optimize compression parameters, transforming a time-consuming manual process into an efficient automated training process.
Data Source
AI summary
An image compression method includes obtaining a hidden variable of an input image using a coding network of a deep learning neural network comprising at least one downsampling back projection module including performing downsampling transformation on a first feature map of the input image input to the downsampling back projection module to obtain a second feature map, obtaining a third feature map having a same resolution as a resolution of the first feature map by reconstructing the second feature map, and obtaining a fourth feature map as an optimization result of the second feature map, based on a difference value between the first feature map and the third feature map; and obtaining a bitstream file of a compressed image by performing entropy coding based on the hidden variable obtained based on the fourth feature map of a last downsampling back projection module.