Variable-Rate Neural Network Video Compression via Image Resizing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based video compression technologies face challenges in achieving variable rate compression without altering network architectures, which limits their flexibility and efficiency in real-world applications.
Innovation Solution
The proposed method involves resizing images to different sizes to achieve variable rate neural network-based compression, allowing for flexible rate adjustments without changing network architectures. This is achieved by using a set of pretrained models and selecting the best candidate based on rate-distortion optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing neural network-based video compression technologies are used, then compression is achieved, but variable rate flexibility is limited due to fixed network architectures
Solution Approach 1:
The patent applies dynamics by making the image size variable rather than fixed. The system dynamically adjusts the input image size to the neural network based on the desired compression rate, allowing the same network architecture to operate at multiple compression rates without retraining. This resolves the contradiction by enabling adaptability through dynamic parameter adjustment rather than through complex architectural changes.
Solution Approach 2:
The patent changes the parameter of image size fed into the neural network to achieve variable rate compression. By modifying this input parameter rather than the network architecture itself, the system can adjust compression rates flexibly while maintaining the same trained model, thus improving adaptability without increasing device complexity.
2Adaptability or versatility
If neural network architectures are altered to achieve variable rate compression, then rate flexibility improves, but model retraining and computational overhead increase
Solution Approach 1:
The patent performs preliminary action by training the neural network once with a fixed architecture for a standard image size. This pre-trained model can then be used for variable rate compression by simply adjusting the input image size parameter, eliminating the need for repeated retraining. This resolves the contradiction by achieving rate flexibility through parameter adjustment rather than through time-consuming architectural modifications and retraining.
3Adaptability or versatility
If image size is resized for variable rate compression, then rate flexibility improves, but processing quality may deteriorate
Solution Approach 1:
The patent makes the neural network multi-functional by designing it to handle multiple compression rates through a single architecture. By training the network to be robust to input size variations and by allowing dynamic size adjustment, the same model serves multiple compression rate requirements, achieving universality. This resolves the contradiction by maintaining processing quality across different rates through a unified, adaptable model rather than through rate-specific models.
Data Source
AI summary
A mechanism for processing video data is disclosed. A determination is made to resize an image with a first size to create a resized image with a second size. A conversion is performed between a visual media data and a bitstream based on the resized image. The conversion includes applying a neural network-based coding model to the resized image to achieve variable rate neural network-based compression.