Video Decoding Neural Network Patch Size Overflow Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing specification for neural network post-filter characteristics in video coding schemes faces issues with patch sizes larger than the picture size and potential overflow of network parameters in integer arithmetic, particularly with 32-bit or 64-bit integers.
Innovation Solution
A video decoding apparatus and coding apparatus are configured to decode and code inverse conversion information using a neural network, ensuring that the patch size information and network parameter values are within valid ranges to prevent overflow and accurately process images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the existing specification for neural network post-filter characteristics is used, then the neural network can be applied for image processing, but the patch size may exceed the picture size causing processing errors
Solution Approach 1:
The patent changes the parameter definition by introducing separate syntax elements for picture width and height, and patch width and height. This allows the patch size parameters to be independently controlled and constrained to be within valid ranges, preventing the patch size from exceeding the picture size while maintaining the neural network processing capability.
2Adaptability or versatility
If the existing specification is used, then neural network parameters can be defined, but the maximum value of network parameters may overflow in 32-bit or 64-bit integer arithmetic
Solution Approach 1:
The patent introduces new syntax elements with explicitly defined bit widths (e.g., 5 bits for log2 picture width, 5 bits for log2 picture height) that constrain the maximum values to prevent overflow in standard integer arithmetic. This maintains parameter definition flexibility while ensuring arithmetic correctness by limiting the range of values to safe boundaries.
3Productivity
If patch size is allowed to be larger than picture size, then the neural network can handle larger processing units, but the decoded image cannot be properly generated
Solution Approach 1:
The patent performs preliminary validation by explicitly decoding and constraining the patch size parameters before the neural network processing occurs. By establishing the relationship that patch width and height must be within valid ranges relative to the picture dimensions, the system prevents invalid processing configurations from occurring, ensuring both efficiency and accuracy.
Data Source
AI summary
A video decoding apparatus includes an image decoding apparatus configured to decode coded data to generate a decoded image, and a resolution inverse conversion apparatus configured to convert a resolution of the decoded image to a specified resolution by using inverse conversion information. The resolution inverse conversion apparatus uses a neural network. In the resolution inverse conversion apparatus, information, indicating a number of horizontal pixels and a number of vertical pixels of a patch size being a unit of processing of the neural network, is decoded. The number of horizontal pixels and the number of vertical pixels of the decoded image are set equal to a maximum value of the number of horizontal pixels and a maximum value of the number of vertical pixels of the patch size, respectively.