Method, computer device, apparatus, and medium for task-adaptive preprocessing
By using the TAPP neural network and neural network encoder-decoder in the TAPP framework and updating the image with RD loss gradient, the problem of neural image compression methods being difficult to adapt across tasks after training is solved, achieving flexible compression task adaptation and model versatility.
CN115461753BActive Publication Date: 2026-07-07TENCENT AMERICA LLC
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TENCENT AMERICA LLC
- Filing Date
- 2021-08-06
- Publication Date
- 2026-07-07
AI Technical Summary
Technical Problem
Existing neural image compression methods are difficult to adapt flexibly to target tasks with different bit rates or quality losses after training, resulting in models that cannot be used across tasks.
Method used
The Task Adaptive Preprocessing (TAPP) framework is adopted, which generates alternative images through TAPP neural networks and combines neural network encoding and decoding. The input image and alternative image are updated using RD loss gradient to adapt to different compression tasks.
Benefits of technology
This study achieves flexible adaptation of neural image compression methods under different target tasks, improving the model's versatility and compression efficiency.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN115461753B_ABST
Abstract
A method for task adaptive pre-processing (TAPP) for neural image compression is performed by at least one processor and includes generating a substitute image based on an input image using a TAPP neural network and encoding the generated substitute image using a first neural network to generate a compressed representation. The TAPP neural network is trained by generating a substitute training image based on an input training image using the TAPP neural network, encoding the generated substitute training image using the first neural network to generate a compressed training representation, decoding the generated compressed training representation using a second neural network to reconstruct an output training image, generating a gradient of a rate-distortion (R-D) loss generated based on the input training image, the reconstructed output training image, and the generated compressed training representation, and updating the generated substitute training image based on the generated gradient of the R-D loss.
Need to check novelty before this filing date? Find Prior Art