Adaptive Subsampling Image Compression via Local JND Profiles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image compression methods, particularly lossy compression, often result in visible artifacts and high data losses, especially in homogeneous image areas, due to the lack of a uniform and explicit visual perception model, leading to suboptimal compression rates and quality.
Innovation Solution
A method involving primary and secondary coding, where data point blocks are processed to determine representative block values and tolerance criteria based on perception models, reducing data loss visibility by adaptively compressing data in the spatial domain, using techniques like JPEG-LS with localized quantization error control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If lossy compression is used to achieve higher compression rates, then compression rate is improved, but visible artifacts and data loss increase
Solution Approach 1:
The patent applies local quality by using Just-Noticeable-Distortion (JND) profiles that are computed locally for different regions of the image. The JND profile adapts to local image characteristics such as texture and luminance, allowing different quantization strengths in different regions. This resolves the contradiction by applying stronger compression in regions where distortions are less noticeable and weaker compression in regions where distortions would be more visible.
Solution Approach 2:
The patent implements dynamics by making the compression parameters adaptive rather than static. The JND profile is dynamically computed based on the local image content, and the quantization step size is adjusted according to the perceived sensitivity of each region. This dynamic adaptation allows the system to achieve higher overall compression rates while maintaining perceptual quality.
2Device complexity
If uniform quantization is applied across the entire image, then device complexity is reduced, but perceptual quality deteriorates due to visible artifacts in sensitive regions
Solution Approach 1:
The patent replaces uniform quantization with local quality-based quantization using JND profiles. Instead of applying a single quantization parameter across the entire image, the system computes local JND values that reflect the perceptual sensitivity of each region. This allows the system to maintain lower complexity while significantly improving perceptual quality by adapting quantization to local image characteristics.
3Productivity
If strong compression is applied to homogeneous areas, then compression rate is improved, but artifacts become more noticeable in those areas
Solution Approach 1:
The patent addresses this contradiction by using JND profiles that specifically account for the perceptual sensitivity of homogeneous regions. The JND computation identifies areas where the human visual system is more sensitive to distortions, such as smooth gradients and homogeneous regions. The system then applies weaker quantization in these sensitive areas and stronger quantization in less sensitive areas, resolving the contradiction between compression rate and artifact visibility.
Data Source
Figure 1~2
Figure 3(a)~3(g)
AI summary
The invention relates to the coding and decoding of data, in particular for visual representation. For this, a method (100) is provided for coding an initial dataset (110), in which a respective initial data value (Ρ1, P2, P3, P4) is established for a respective plurality of initial data points (112), for generating a compressed dataset (114), comprising a primary coding of the initial dataset (110) to generate an intermediate dataset (116) and a secondary coding of the intermediate dataset (116) to generate the compressed dataset (114). In this way, the primary coding comprises the following steps: establishing a plurality of data point blocks (118), each having a plurality of initial data points (112) of the initial dataset (110); determining a block data value (120) for each data point block (118) from the plurality of initial data values (P1, P2, P3, P4) within the respective data point block (118); checking each data point block (118) for compliance with a tolerance criterium; and generating the intermediate dataset (116) in such a way that the intermediate dataset (116) contains the block data value (120) as an individual data value for each data point block (118) in which the tolerance criterium is observed, and same contains the individual initial data values (P1, P2, P3, P4) at the initial data points (112) of the respective data point block (118) for each data point block (118) in which the tolerance criterium is not observed.