Neural Video Coding with Padding and Cropping for Tensor Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression techniques face challenges in achieving high compression ratios with minimal loss of picture quality, particularly when dealing with limited network resources and varying signal dimensions.
Innovation Solution
Implementing a neural network framework with multi-stage context models and tensor boundary handling, including padding and cropping layers to manage non-integer tensor sizes, ensuring device interoperability and reducing memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If video data is compressed to reduce data size for transmission or storage, then the quantity of data is decreased, but picture quality may be sacrificed
Solution Approach 1:
The video data is divided into multiple blocks that are processed independently through the neural network. Each block undergoes separate encoding and decoding operations, allowing for optimized compression while maintaining overall picture quality through distributed processing across multiple segments
Solution Approach 2:
The neural network dynamically adjusts encoding parameters such as quantization levels, transformation coefficients, and prediction modes based on the complexity and characteristics of each video block. This adaptive parameter adjustment enables higher compression ratios for simple regions while preserving quality for complex regions
2Productivity
If neural networks are applied to encode and decode video data, then coding efficiency is improved, but device interoperability may be compromised due to non-integer tensor sizes
Solution Approach 1:
Padding layers are applied to video blocks before they enter the neural network processing stages. This preliminary action ensures that all tensors have integer dimensions compatible with standard device architectures, preventing interoperability issues while allowing the network to achieve optimal coding efficiency through subsequent processing
Solution Approach 2:
The padding and cropping layers act as intermediary components between the variable-sized video blocks and the fixed-architecture neural network. These intermediaries transform the input data into a format that maintains both the information content and the integer dimension requirements for cross-device compatibility
3Reliability
If padding and cropping layers are added to handle tensor boundaries, then device interoperability is ensured, but the complexity of the neural network framework increases
Solution Approach 1:
The padding and cropping layers are designed as universal components that can be applied to video blocks of any size and configuration. These multi-functional layers handle both boundary management and dimension normalization in a single integrated framework, reducing overall system complexity while ensuring device interoperability across different platforms
Data Source
AI summary
The present disclosure pertains to methods, neural networks, encoders and decoders for processing a picture. Specifically, padding layers are added preceding the layers that has same function as down-sampling layers and cropping layers are added following the layers that has same function as up-sampling layers so as to reduce amount of the data processed in the neural network and improve coding efficiency.