Multi-Scale Context Model for Image Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image encoding methods using single-scale context models fail to achieve rate-distortion balance, resulting in the need for more bits to denote image data and suboptimal compression efficiency.
Innovation Solution
The implementation of a multi-scale context model with multiple mask convolution layers of different scales to estimate probabilities of symbols, followed by entropy encoding, which improves the accuracy of probability estimation and reduces the number of bits required to represent image data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single-scale context model is used to estimate symbol probabilities, then the encoding process is simpler, but the probability estimation accuracy is insufficient leading to higher bit rates
Solution Approach 1:
The patent divides the context modeling process into multiple scales (e.g., 3x3, 5x5, 7x7 convolutional kernels) that operate independently to estimate probabilities at different spatial contexts. Each scale captures local dependencies separately, and their outputs are combined to form the final probability estimate, thereby improving accuracy without overwhelming complexity
Solution Approach 2:
The patent transitions from single-scale to multi-scale context modeling by adding the dimension of scale variation. Different convolutional kernel sizes (3x3, 5x5, 7x7) represent different spatial extents, creating a hierarchical structure that captures context information at multiple levels of abstraction simultaneously
2Loss of information
If more bits are used to denote image data, then the image quality is better preserved, but the compression efficiency deteriorates
Solution Approach 1:
The patent changes the parameter of context window size by employing multiple scales (3x3, 5x5, 7x7 kernels). This allows the system to adaptively capture context information at different spatial extents, leading to more accurate probability estimates and consequently more efficient entropy encoding that achieves better compression while preserving image fidelity
3Productivity
If the probability estimation is inaccurate, then the encoding process is faster, but the number of bits required increases
Solution Approach 1:
The patent performs preliminary probability estimation using multi-scale context models before the actual entropy encoding process. By pre-computing accurate probability distributions from multiple scales and combining them, the system prepares optimal encoding parameters in advance, enabling faster encoding without sacrificing compression efficiency
Data Source
AI summary
Embodiments of this disclosure provide an image encoding method and apparatus and image decoding method and apparatus. The image encoding includes performing convolutional neural network (CNN) encoding on image data to generate feature vectors or feature maps; quantizing the feature vectors or feature maps to generate discrete symbols to be encoded; and estimating probabilities of the symbols to be encoded by using a multi-scale context model including multiple mask convolution layers of different scales. An entropy encoding of the image data is performed according to the probabilities of the symbols to be encoded.