Multi-Resolution Visual Search via Hierarchical Feature Descriptors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visual search systems face challenges in accurately classifying objects from images with perspective deformations due to the linear scaling of databases with the number of descriptors and images, leading to errors in dictionary representation and high variance in tree classifiers, which results in misclassification and reduced robustness against viewpoint variations.
Innovation Solution
The implementation of a hierarchical data structure using multi-resolution feature descriptors and a normalized bagging kernel to improve robustness against perspective deformations and high variance issues, by generating a hierarchically arranged data structure for classifying objects and applying a partitioning algorithm to determine classifying data structures, and using bagging to reduce variance in classifiers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a linear scaling database approach is used to store and search feature descriptors, then the system can handle large numbers of images and descriptors, but the classification accuracy deteriorates due to errors in dictionary representation and high variance in tree classifiers
Solution Approach 1:
The patent divides the large database into multiple subsets, each processed by a separate tree classifier. Instead of building one large classifier that degrades in accuracy, the system segments the data and uses multiple smaller classifiers (forest) to maintain precision while handling large quantities of images and descriptors
Solution Approach 2:
The patent applies preliminary actions by extracting multi-resolution feature descriptors before classification, and by pre-processing images to detect and correct perspective deformations. This preliminary processing improves the quality of input data for the classifiers, thereby maintaining accuracy even with large datasets
2Device complexity
If a single tree classifier is used for visual search, then the system structure remains simple, but the variance in classification results increases leading to misclassification
Solution Approach 1:
The patent merges multiple tree classifiers into a random forest ensemble. By combining the results of multiple classifiers that each process different subsets of data, the system reduces variance and improves classification consistency while maintaining a relatively simple overall structure
Solution Approach 2:
The patent changes parameters by using multi-resolution feature descriptors (varying the resolution parameter) and by introducing randomness in data sampling and feature selection for each tree in the forest. These parameter changes enable the system to handle complexity while improving reliability through diverse classification perspectives
3Speed
If feature descriptors are extracted without considering multi-resolution, then the processing speed remains fast, but the robustness against perspective deformations decreases
Solution Approach 1:
The patent applies partial action by extracting feature descriptors at multiple resolutions rather than a single resolution. This partial multi-resolution approach provides sufficient robustness against perspective deformations while avoiding the excessive computational cost of full multi-resolution analysis, thus maintaining acceptable processing speed
Solution Approach 2:
The patent adds the resolution dimension to feature descriptor extraction. By extracting descriptors at multiple scales (adding a resolution dimension), the system gains robustness to perspective deformations without significantly impacting speed, as the additional dimensional information is efficiently processed
Data Source
AI summary
In general, techniques are described for performing a vocabulary-based visual search using multi-resolution feature descriptors. A device may comprise one or more processors configured to perform the techniques. The processors may generate a hierarchically arranged data structure to be used when classifying objects included within a query image based on multi-resolution query feature descriptor extracted from the query image at a first scale space resolution and a second scale space resolution. The hierarchically arranged data structure may represent a first query feature descriptor of the multi-resolution feature descriptor extracted at the first scale space resolution and a second corresponding query feature descriptor of the multi-resolution feature descriptor extracted at the second scale space resolution hierarchically arranged according to the first scale space resolution and the second scale space resolution. The processors may then perform a visual search based on the generated data structure.