Image Search Accuracy via Multi-Granularity Mining Network

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image search technologies face challenges in accurately and efficiently searching for images with similar objects due to the inability to distinguish between background and object features, leading to inaccurate feature extraction and reduced search results.

Innovation Solution

The implementation of a multi-granularity mining network (MGMN) with an adaptive spatial attention unit (ASAU) that employs deep learning to highlight target areas, filter out background information, and combine features from different scales, enhancing the accuracy of image search results.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep learning-based AI systems are used to improve recognition rates and understanding of user preferences, then the accuracy of image feature extraction is improved, but the computational complexity and processing time increase

Engineering Contradiction:
Improveimage feature extraction accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the image processing task into distinct functional modules: a backbone network for extracting basic feature information, and a feature representation network for processing and refining these features. This segmentation allows each module to specialize in specific operations, improving overall accuracy while enabling independent optimization and parallel processing to mitigate computational complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The backbone network performs preliminary feature extraction on the input image before the feature representation network processes it further. By pre-processing the image to extract basic features first, the system reduces the computational burden on subsequent stages while maintaining high accuracy in the final feature representation.

Inventive Principle:
Principle #10Preliminary action

2Loss of information

If feature extraction includes both background and object information, then comprehensive image representation is achieved, but the precision of target object identification decreases

Engineering Contradiction:
Improveimage information completenessVSAvoidtarget object identification accuracy
Core Design Contradiction:
Loss of informationVSMeasurement precision

Solution Approach 1:

The feature representation network extracts and isolates the target object features from the comprehensive feature information generated by the backbone network. By separating the target object information from background and other irrelevant information, the system achieves precise target identification while retaining the option to use comprehensive features when needed.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different processing qualities to different parts of the feature information. The feature representation network focuses computational resources on enhancing the quality of target object features while applying lighter processing to background information, thereby improving target identification accuracy without completely discarding comprehensive image representation.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3908943B1Method, apparatus, electronic device and computer readable storage medium for image searching
Publication Date: 2024.10.09 SAMSUNG ELECTRONICS CO LTD
  • EP3908943B1 patent drawingFigure 1~3
  • EP3908943B1 patent drawingFigure 4a
  • EP3908943B1 patent drawingFigure 4b

AI summary

A method and apparatus for image searching based on artificial intelligent (AI) are provided. The method includes obtaining first feature information by extracting features from an image based on a first neural network, obtaining second feature information corresponding to a target area of a query image by processing the first feature information based on a second neural network and at least two filters having different sizes, and identifying an image corresponding to the query image according to the second feature information.