Frequency Domain Interest Point Descriptor for Image Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image feature descriptors, such as SIFT and SURF, are computationally expensive and have large memory footprints due to their reliance on floating-point operations and large vector sizes, making them inefficient for large-scale image database matching.
Innovation Solution
A compact interest point descriptor is created by dividing an image area into subareas, transforming them into the frequency domain using DCT, normalizing for contrast invariance, and representing them as bitwise strings, with distances calculated using Hamming distance for efficient matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If SIFT and SURF descriptors are used for image feature description, then distinctive mathematical description is provided, but memory consumption increases significantly due to large descriptor sizes (144-512 bytes)
Solution Approach 1:
The patent divides the image area into multiple subareas and processes each subarea independently to create separate bit strings, which are then concatenated. This segmentation approach allows the descriptor to maintain distinctive information from different regions while using compact bitwise representation instead of large floating-point vectors.
Solution Approach 2:
The patent transforms the feature representation from floating-point vectors (used in SIFT/SURF) to binary bit strings through frequency domain analysis and thresholding. This parameter change from continuous to discrete representation dramatically reduces memory footprint while preserving discriminative power through frequency-based feature selection.
2Measurement precision
If SIFT and SURF descriptors use floating point vectors for distance computation, then accurate similarity measurement is achieved, but computational time increases significantly
Solution Approach 1:
The patent replaces the mechanical computation of Euclidean distances between floating-point vectors with bitwise XOR operations and Hamming distance calculations on binary strings. This substitution leverages efficient bitwise operations native to computer architecture, dramatically reducing computational time while maintaining measurement capability through frequency domain feature selection.
3Productivity
If frequency domain representation with bitwise operations is used, then matching speed increases significantly, but descriptor creation complexity increases due to DFT/DCT transformation
Solution Approach 1:
The patent extracts only the essential frequency components from the full frequency domain representation by selecting specific frequencies and comparing them to thresholds. This extraction approach creates compact binary descriptors that capture the most discriminative information while avoiding the complexity of representing or processing the entire frequency spectrum.
Data Source
AI summary
Systems and methods for image analysis and recognition are disclosed, in particular the methods for interest point description. An interest point and its surrounding area is broken into subareas, a frequency domain description of each area is created by applying discrete Fourier transform (DFT). Frequency domain features are than coded bitwise by comparing them to predefined thresholds. Subsequently, the present invention provides alternative or improved methods and data structures for interest point description that may reduce memory consumption and allow fast bitwise matching.