Digital Image Hash Clustering for Faster Similarity Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional digital image search techniques face significant challenges due to the ever-increasing number of digital images, leading to excessive computational and power resource consumption, particularly in service provider systems like stock image and social media platforms, as they require linear comparisons of digital image hashes, resulting in a more than linear increase in computational complexity.
Innovation Solution
The implementation of digital image hash search techniques that form clusters based on cluster centers and use a distance measure to identify relevant clusters, allowing for a two-stage search process where comparisons are terminated once a distance threshold is reached, thereby reducing unnecessary computations and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional linear search techniques are used to compare digital image hashes, then search completeness is maintained, but computational resource consumption increases more than linearly with the number of images
Solution Approach 1:
The patent segments the large set of digital image hashes into multiple clusters, each represented by a cluster center. Instead of comparing the query hash against all individual image hashes, the system first identifies relevant clusters based on distance to cluster centers, then performs detailed comparisons only within those clusters. This segmentation reduces the search space from O(N) to O(k + m) where k is the number of clusters and m is the average cluster size, significantly improving search speed while reducing computational resource consumption.
Solution Approach 2:
The patent introduces cluster centers as intermediary representations that mediate between the query hash and individual image hashes. The cluster center serves as a representative proxy for all images in its cluster, allowing the system to quickly eliminate entire clusters that are unlikely to contain matches. This intermediary layer reduces the number of direct hash comparisons needed, thereby reducing computational resource consumption while maintaining search completeness.
2Measurement precision
If all digital image hashes are compared to ensure complete search results, then search accuracy is maintained, but computational time increases significantly
Solution Approach 1:
The patent performs preliminary clustering of digital image hashes before the actual search operation. By pre-organizing images into clusters based on their hash similarities and computing cluster centers in advance, the system creates a structured search space that enables faster query processing. This preliminary action ensures that during search, the system can quickly identify relevant clusters and perform accurate comparisons only within those clusters, maintaining search accuracy while reducing computational time.
Solution Approach 2:
The patent applies a distance threshold to limit comparisons to only those images within a certain similarity radius of the query hash. By using the cluster center distance as a filter, the system performs partial comparisons - only within identified clusters - rather than exhaustive comparisons across all images. This partial action maintains search accuracy for relevant results while significantly reducing computational time by excluding clearly unrelated images.
Data Source
AI summary
Digital image hash search techniques are described. These techniques leverage clusters formed from digital image hashes that overcome limitations and computational resource consumption of conventional clustering techniques used to implement a digital image search. In an example, search techniques employ two search stages. In a first stage, clusters are identified based on the cluster centers using a distance measure. The second stage involves a comparison of cluster hashes within the identified cluster with the search query hash until the distance measure is reached.