Image Encoding Using Asymmetric Probability Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image encoding methods waste bits by assuming a symmetrical prediction residual distribution, leading to unnecessary encoding of the right half of the distribution, which is never used, and fail to utilize the predicted value effectively, resulting in inefficient compression.
Innovation Solution
Encode the original pixel value using a conditional distribution that considers the upper and lower limits of possible pixel values, utilizing the predicted value to determine the probability distribution of the original pixel value, thereby avoiding the computation of prediction residuals and optimizing encoding efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of substance
If prediction residual distribution is assumed to be symmetrical, then encoding can be performed using conventional methods, but unnecessary bits are wasted by encoding the right half of the distribution that is never used
Solution Approach 1:
The patent applies asymmetry by recognizing that the prediction residual distribution is actually asymmetrical (concentrated at values near 0 with bounded range), not symmetrical as conventionally assumed. By designing the encoding scheme to match this asymmetrical distribution - using asymmetric probability models and asymmetric code allocation - the patent eliminates waste of bits on impossible values while maintaining decoding simplicity.
2Productivity
If predicted value is not fully utilized, then encoding can be performed independently of prediction accuracy, but encoding efficiency is reduced due to unnecessary computation of prediction residuals
Solution Approach 1:
The patent applies preliminary action by fully utilizing the predicted value during the encoding process. Instead of computing prediction residuals (x - x') and then encoding those residuals, the method directly encodes the original pixel value x using the predicted value x' to guide the probability distribution model. This preliminary utilization of prediction information eliminates unnecessary residual computation and improves encoding efficiency.
Solution Approach 2:
The patent substitutes the mechanical computation of prediction residuals with a more efficient information-theoretic approach. Rather than performing arithmetic subtraction and then encoding the result, the system uses the predicted value to directly shape the probability distribution for entropy encoding of the original value, replacing the residual computation mechanism with a distribution-based encoding mechanism.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
An image encoding method for encoding a pixel value of an encoding target by using a predicted value generated by means of spatial or temporal prediction using a previously-decoded image. The method performs prediction of the pixel value of the encoding target and obtains the predicted value; computes data of a probability distribution which indicates what value an original pixel value has for the obtained predicted value, by shifting, in accordance with the predicted value, difference distribution data of a difference between the original pixel value and the predicted value in predictive encoding, where the difference distribution data is stored in advance; clips the obtained data of the probability distribution so as to contain the data in a range from a lower limit to an upper limit for possible values of the original pixel value; and encodes the pixel value of the encoding target by using the clipped data of the probability distribution of the original pixel value from the lower limit to the upper limit.