Image Retrieval Hash Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing image retrieval technologies face inefficiencies due to the complexity and time-consuming nature of calculating similarity between images, especially when dealing with large databases, leading to low retrieval efficiency.

Innovation Solution

The method involves obtaining a hash value of a reference image, splitting it into segments, and using these segments to retrieve matching hash value segments from target hash tables, allowing for the identification of candidate images with similar features by calculating hamming distances only between these candidates and the reference image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If similarity calculation is performed between the reference image and each image in the database, then comprehensive image retrieval is achieved, but the data processing amount becomes excessively large and retrieval efficiency becomes excessively low

Engineering Contradiction:
Improvecomprehensive image retrievalVSAvoidretrieval efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent divides the hash value of each image into multiple segments (e.g., 4 segments of 16 bits each from a 64-bit hash value). Instead of comparing complete hash values, the system creates separate hash tables for each segment position and compares only corresponding segments. This segmentation reduces the comparison scope from N images to a subset of images sharing at least one matching segment, significantly improving retrieval efficiency while maintaining comprehensive search capability.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If similarity calculation is performed between the reference image and each image in the database, then all candidate images are evaluated, but the calculation complexity and time consumption become excessively high

Engineering Contradiction:
Improvesimilarity calculation accuracyVSAvoidcalculation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent pre-divides hash values into segments and organizes them into multiple hash tables before retrieval operations. During retrieval, the system first performs quick segment matching to identify candidate images, then calculates full similarity only for these candidates. This preliminary segmentation and indexing action reduces the complexity of subsequent similarity calculations by filtering out non-matching images beforehand.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If the database contains a large quantity of images, then comprehensive coverage is achieved, but the calculation amount for similarity comparison becomes excessively large

Engineering Contradiction:
Improvedatabase image quantityVSAvoiddata processing amount
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent transforms the one-dimensional problem of comparing complete hash values into a multi-dimensional approach by creating multiple hash tables, each corresponding to a specific segment position. This dimensional transformation allows the system to query along different segment dimensions and intersect results, efficiently handling large databases by reducing the search space from O(N) to O(N/k) where k is the number of segments.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach significantly reduces the data processing amount and improves image retrieval efficiency by focusing on a smaller set of candidate images with matching hash value segments, thereby speeding up the retrieval process.

Implementation Method 1

calculating a hamming distance between each of the candidate images and the reference image according to the first hash value of the reference image and the second hash values of the candidate images; and retrieving, from the candidate images to which the second hash value segments respectively belong, the target image whose hamming distance to the reference image is less than a preset threshold

Methodology Applied
Scientific EffectHamming distance:

Data Source

PatentEP3767483B1Method, device, system, and server for image retrieval, and storage medium
Publication Date: 2024.09.18 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3767483B1 patent drawingFigure 1
  • EP3767483B1 patent drawingFigure 2a
  • EP3767483B1 patent drawingFigure 2b

AI summary

A method, device, system, and server for image retrieval, and a storage medium. The method comprises: acquiring first hash value of a reference image to be retrieved; splitting the first hash value into a preset number of first hash value segments, where the different first hash value segments are different in terms of positional order in the first hash value; positioning a target hash table on the basis of the positional order of the first hash value segments, retrieving from multiple hash value segments in the target hash table second hash value segments identical to the first hash value segments, the target hash table being used for storing hash value segments found on the positional order of the first hash value segments in hash values of images; and retrieving a target image similar to the reference image from candidate images to which the second hash value segments belong. The method reduces the data processing volume in an image retrieval process and increases the efficiency of image retrieval.