Object Recognition Using Extended Inverted File Structure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object recognition systems face challenges in scalability and efficiency when dealing with large collections of images, particularly in recognizing small objects buried in complex backgrounds, and are not well-suited for real-time identification of multiple objects due to reliance on TF-IDF scoring and lack of spatial consistency validation.
Innovation Solution
A method that performs object recognition in a single step by using the vocabulary of visual words and an extended inverted file structure, allowing for direct validation of spatial consistency and efficient indexing of reference images, enabling rapid identification of objects in query images regardless of location, scale, or orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If TF-IDF scoring and traditional multi-step recognition methods are used, then recognition robustness is maintained, but recognition speed and scalability to large collections deteriorate
Solution Approach 1:
The patent segments the recognition process into two distinct phases: an offline indexing phase where reference images are pre-processed and organized into an extended inverted file structure, and an online recognition phase where query images are rapidly matched against the pre-built index. This segmentation allows computationally intensive operations to be performed offline, enabling fast real-time recognition without sacrificing robustness.
Solution Approach 2:
The patent performs preliminary actions by pre-computing and storing visual word frequencies, spatial distribution statistics, and inverted file structures for all reference images before actual recognition occurs. This preliminary indexing includes calculating term frequency-inverse document frequency (TF-IDF) scores and organizing spatial consistency data, so that during online recognition, only lightweight matching operations are needed.
2Device complexity
If traditional indexing methods are used, then system complexity is low, but scalability to large collections of images deteriorates
Solution Approach 1:
The patent extends the traditional inverted file structure by adding spatial dimension information to the indexing process. Instead of only indexing visual words, the extended inverted file structure indexes visual words along with their spatial coordinates, frequency distributions, and consistency metrics. This dimensional extension enables the system to scale to large collections while maintaining the relatively simple inverted file structure paradigm.
3Productivity
If spatial consistency validation is omitted, then recognition speed increases, but recognition accuracy for small objects in complex backgrounds deteriorates
Solution Approach 1:
The patent introduces spatial consistency metrics as an intermediary validation layer between visual word matching and final recognition decisions. Rather than performing full geometric verification on all candidate matches, the system uses pre-computed spatial distribution statistics and frequency-based metrics to quickly filter implausible matches, maintaining high speed while improving accuracy for small objects in complex backgrounds.
4Measurement precision
If multiple-step recognition processes are used, then recognition accuracy is maintained, but processing time increases
Solution Approach 1:
The patent merges multiple traditional recognition steps into a single integrated rapid recognition process. By combining visual word matching, spatial consistency checking, and scoring into one unified operation that leverages the pre-built extended inverted file structure, the system achieves recognition accuracy comparable to multi-step methods while reducing processing time to under five seconds for large collections.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Identification of objects in images. All images are scanned for key- points and a descriptor is computed for each region. A large number of descriptor examples are clustered into a Vocabulary of Visual Words. An inverted file structure is extended to support clustering of matches in the pose space. It has a hit list for every visual word, which stores all occurrences of the word in all reference images. Every hit stores an identifier of the reference image where the key-point was detected and its scale and orientation. Recognition starts by assigning key-points from the query image to the closest visual words. Then, every pairing of the key-point and one of the hits from the list casts a vote into a pose accumulator corresponding to the reference image where the hit was found. Every pair key-point/hit predicts specific orientation and scale of the model represented by the reference image.