Feature Descriptor Compression Using Non-Uniform Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for compressing feature descriptors, such as SIFT, result in large data sizes due to the 128-byte vectors, and existing solutions either omit bits or reduce the number of vectors, which may compromise representation accuracy.
Innovation Solution
A method and device that non-uniformly quantize the feature descriptor and apply run-length coding to reduce the data size, utilizing a non-uniform quantizer and run-length encoder to generate a compressed feature descriptor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the feature descriptor uses 128 vectors with 1 byte each, then the representation accuracy is maintained, but the data size becomes large (128 bytes)
Solution Approach 1:
The patent applies non-uniform quantization to transform the feature descriptor vectors from uniform 1-byte representation to a variable-length representation. By changing the parameter representation method (from fixed-width to variable-width based on magnitude), the patent achieves both compression and accuracy preservation. Vectors with small magnitudes are quantized to zero or small values, while significant vectors retain their full precision, thus reducing overall data size while maintaining representation accuracy for important features.
Solution Approach 2:
The patent extracts and processes significant vectors separately from insignificant ones through non-uniform quantization. By identifying vectors with magnitudes below a threshold and quantizing them to zero or simplified representations, the patent separates the essential information (significant vectors) from redundant information (insignificant vectors), thereby reducing data size while preserving the critical representation accuracy.
2Quantity of substance
If existing methods omit bits or reduce the number of vectors, then the data size is reduced, but the representation accuracy is compromised
Solution Approach 1:
Instead of uniformly reducing all vectors or omitting bits, the patent changes the parameter representation by applying non-uniform quantization. This allows the system to adaptively allocate precision: vectors with small magnitudes (less important for representation) are compressed more aggressively, while vectors with large magnitudes (critical for accuracy) retain their full precision. This selective parameter transformation resolves the contradiction by making the compression adaptive rather than uniform.
Solution Approach 2:
The patent applies different compression strategies to different vectors based on their local characteristics (magnitude). Significant vectors (with large magnitudes) receive minimal compression to preserve accuracy, while insignificant vectors (with small magnitudes) undergo aggressive quantization. This local differentiation in quality preservation allows the system to reduce overall data size while maintaining representation accuracy where it matters most.
3Quantity of substance
If uniform quantization is applied to reduce data size, then compression is achieved, but the compression ratio is insufficient
Solution Approach 1:
The patent transforms the quantization parameter from a uniform fixed-step approach to a non-uniform adaptive approach. By changing the quantization step size based on the vector magnitude and significance, the patent achieves higher compression ratios. Significant vectors use larger quantization steps (preserving accuracy), while insignificant vectors use smaller or zero steps (achieving higher compression), thereby improving overall compression ratio beyond what uniform quantization can achieve.
Solution Approach 2:
The patent introduces dynamic adaptivity into the quantization process, where the quantization parameters are adjusted based on the actual content of each vector rather than applying a static uniform approach. This dynamic adjustment allows the compression algorithm to respond to the actual data characteristics, achieving higher compression ratios by being more aggressive with insignificant vectors while maintaining accuracy for significant ones.
Data Source
AI summary
A device for compressing a feature descriptor includes a non-uniform quantizer and a run-length encoder. The non-uniform quantizer accesses a source feature descriptor from a storage device, and non-uniformly quantizes the source feature descriptor having source vectors into an intermediate feature descriptor having intermediate vectors according to a vector default value. The run-length encoder executes run-length coding for the intermediate feature descriptor to generate a compressed feature descriptor.