Image Database Clustering for Object Recognition Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The processing time and memory requirements for comparing and storing local features in large image databases are excessive, making it inefficient for specific object recognition, especially when dealing with large numbers of images.
Innovation Solution
A method that clusters local feature vectors and selects representative vectors based on their length, storing these representatives in the database for efficient comparison, reducing memory usage without significantly compromising recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all local feature vectors are stored in the image database, then recognition accuracy is maintained, but memory usage becomes excessive
Solution Approach 1:
The patent extracts and stores only representative feature vectors from clusters of similar local features, rather than storing all feature vectors. This selective extraction maintains recognition accuracy while significantly reducing memory usage by eliminating redundant storage of highly similar feature vectors.
Solution Approach 2:
The patent merges highly similar feature vectors into clusters and represents each cluster by a single representative feature vector. This merging process reduces the total number of stored vectors while preserving the essential visual information needed for accurate object recognition.
2Reliability
If multiple local features are extracted for high accuracy recognition, then recognition robustness improves, but processing time increases
Solution Approach 1:
The patent performs clustering and selection of representative feature vectors in advance during database compilation. This preliminary action organizes the feature vectors into clusters and identifies representatives beforehand, so that during actual recognition, the system only needs to compare query vectors with pre-selected representatives, significantly reducing processing time while maintaining robustness.
3Quantity of substance
If feature vectors are reduced for memory efficiency, then storage requirements decrease, but recognition rate may decrease
Solution Approach 1:
The patent applies different treatment to different feature vectors based on their local characteristics. Representative feature vectors that best represent their clusters are selected for storage, while redundant vectors are discarded. This local quality approach ensures that the most informative features are retained, maintaining high recognition rates with reduced storage requirements.
Data Source
AI summary
A method for creating an image database comprising an extraction step of extracting reference feature vectors from a reference image which should be compared with a retrieval query image for object recognition, the reference feature vectors corresponding to local features at different positions of the reference image and representing the position and characteristics of each of the local features as a vector position, vector length, and a vector direction, a clustering step of creating a plurality of clusters consisting of different reference feature vectors in such a manner that each reference vector belongs to any of the plurality of clusters, a selection step of selecting the representative vector of the clusters from among the reference feature vectors of each of the clusters, and a step of associating the representative vector with the reference image and registering the representative vector associated therewith in the image database for the object recognition, wherein, in the clustering step, each of the clusters is created in such a manner that reference feature vectors at a near vector position belong to the same cluster, and in the selection step, reference feature vectors with long vector length are given priority to select the representative vector, and wherein the retrieval query image and the reference image are compared with each other by generating at least one query feature vector from the retrieval query image, and applying local search between the query feature vector and the representative vector, each of the steps being executed by computers.