Image Target Search Using Tree Classification Index

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image search methods lack efficiency and accuracy due to the lack of distinction between different target types, leading to irrelevant comparisons and resource wastage, as they primarily focus on local features like SIFT and SURF, ignoring global features such as color, shape, and texture, and do not classify images according to target types.

Innovation Solution

The method involves at least two levels of classification for the target, generating a local feature code, and searching within the classified category, using techniques like deep learning classifiers and tree classification index structures to enhance accuracy and efficiency by matching the target within its specific classification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If targets are compared one by one without classification, then all possible targets are searched, but searching efficiency is low and resources are wasted

Engineering Contradiction:
Improvesearching efficiencyVSAvoidresource wastage
Core Design Contradiction:
ProductivityVSLoss of energy

Solution Approach 1:

The patent segments the search space by dividing targets into different classification categories (e.g., vehicle types, object types). Instead of comparing the query target with all images in the database, the system first classifies the query target, then only searches within the corresponding classification category. This segmentation dramatically reduces the number of comparisons needed, improving searching efficiency while reducing computational resource wastage.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If only local features are used for image matching, then local details are captured, but global features like color, shape, and texture are ignored resulting in lower search accuracy

Engineering Contradiction:
Improvesearch accuracyVSAvoidglobal feature information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent merges multiple feature types for comprehensive target description. It combines local features (SIFT, SURF) with global features including color histograms, shape descriptors, and texture features. By integrating these diverse feature representations, the system achieves more accurate target matching while preserving both local detail information and global contextual information, thereby improving search accuracy without losing important visual characteristics.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If all target types are searched without distinction, then comprehensive coverage is achieved, but relevance of search results decreases

Engineering Contradiction:
Improvesearch coverageVSAvoidresult relevance
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent performs preliminary classification of the query target before conducting the actual search. By first determining the classification category (e.g., identifying it as a 'vehicle' or 'animal'), the system can then limit the search to only relevant categories in the database. This preliminary action maintains comprehensive coverage within the relevant domain while significantly improving result relevance by excluding unrelated target types from the search results.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3327583B1Method and device for searching a target in an image
Publication Date: 2023.03.01 HANGZHOU HIKVISION DIGITAL TECHNOLOGY CO LTD
  • EP3327583B1 patent drawingFigure 1
  • EP3327583B1 patent drawingFigure 2
  • EP3327583B1 patent drawingFigure 3

AI summary

The present application discloses a method and a device for searching a target in an image. The search method includes: locating a target in an image to be searched (S101); classifying the target and extracting a local feature of the target to generate a local feature code of the local feature (S102); searching an image database for suspected images of the same classification as the target through a tree classification index structure, comparing pre-stored local feature codes of the suspected images with the local feature code of the target, and outputting a suspected image with a similarity larger than a first threshold (S103). By locating a target in an image to be searched, performing at least two levels of classification on the target, generating local feature code of the target, and searching according to the classification of the target when searching and sorting by similarity, thus the target to be searched is model-matched within the same target classification, enhancing the searching accuracy and efficiency.