Neural Image Decoding With Linear Correction After Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding and decoding technologies using artificial intelligence (AI) face challenges in effectively addressing coding errors and inefficiencies due to quantization, leading to performance degradation over time, especially in frames distant from intra frames, and error propagation.
Innovation Solution
Implementing a neural network-based approach that includes linear correction parameters derived at the encoder side and applied at the decoder side to correct errors, using techniques like linear regression and backpropagation to optimize image reconstruction, thereby improving coding efficiency and image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If neural network-based encoding is used to improve image quality, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The neural network-based image encoding system is divided into separate functional modules: an encoding neural network for feature extraction, a quantization module for compression, and a decoding neural network for reconstruction. This segmentation allows each component to be optimized independently while maintaining overall system performance, reducing the complexity burden on any single module.
Solution Approach 2:
The encoding neural network performs preliminary feature extraction and transformation before quantization, converting the original image into a compressed feature representation. This preliminary action enables more efficient subsequent processing and reduces the complexity of the decoding stage by pre-organizing the data in a neural network-compatible format.
2Loss of substance
If quantization is applied to compress feature data, then loss of substance is reduced, but manufacturing precision deteriorates
Solution Approach 1:
The quantization process transforms continuous feature values into discrete representations by changing the parameter precision. The patent employs adaptive quantization strategies that adjust the quantization step size based on feature importance, preserving critical information while achieving compression. This parameter transformation enables lossy compression with controlled quality degradation.
Solution Approach 2:
The system incorporates feedback mechanisms where the decoded image quality is evaluated and used to adjust the quantization parameters for subsequent encoding. This feedback loop allows the system to learn from quantization errors and optimize the balance between compression ratio and image quality over time, mitigating the precision loss from quantization.
3Reliability
If linear correction parameters are applied to correct decoding errors, then reliability is improved, but device complexity increases
Solution Approach 1:
Linear correction parameters serve as an intermediary mechanism between the encoded features and the final decoded image. These parameters are derived from the encoding process and applied during decoding to compensate for quantization errors and neural network approximation errors. This intermediary correction layer improves reliability without requiring complete re-encoding, adding minimal complexity to the decoding pipeline.
Solution Approach 2:
The linear correction parameters represent a transformation of the decoding process parameters, introducing simple linear adjustments (scaling and offset) to the decoded features. This parameter change approach provides a computationally efficient error correction mechanism that significantly improves reliability with minimal increase in device complexity, as linear operations are much less demanding than full neural network inference.
Data Source
AI summary
An image decoding method includes obtaining, from a bitstream, feature data obtained via neural network-based encoding of a current image, and linear correction parameters for the current image, obtaining image data for the current image by inputting the feature data to a decoding neural network, and reconstructing the current image by applying the linear correction parameters to the image data.