AI Image Search Using Object Detection to Fix Text Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional text-based image search technologies face accuracy issues due to incorrect image descriptions and biases in search results, requiring significant resources for ranking signal development and customization, and are limited by errors in matching operations.
Innovation Solution
A method and apparatus for image searching using artificial intelligence (AI) that combines natural language processing and object detection to match user queries with image search results, generating keyword-category combinations and storing them as cache data for improved accuracy and relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If text-based image search technology is used, then search engine can operate with existing indexing methods, but search accuracy deteriorates when image text descriptions are incorrect
Solution Approach 1:
The patent replaces the mechanical text-based indexing system with an AI-based object detection system. Instead of relying on text descriptions attached to images, the system uses deep learning models to automatically detect and recognize objects in images, thereby eliminating the dependency on potentially incorrect text annotations while maintaining ease of operation through automated processing
Solution Approach 2:
The patent introduces an intermediary AI processing layer between the user query and the image database. This intermediary component (the AI object detection system) acts as a mediator that translates text queries into object category representations and compares them with detected objects in images, thereby improving search accuracy without requiring direct manipulation of image files or text annotations
2Productivity
If text-based image search with category codes is used, then search engine can index images, but search results show bias depending on the number of labeled images for each query
Solution Approach 1:
The patent substitutes the manual category coding system with an automated AI object detection system. This replacement eliminates the bias introduced by uneven manual labeling efforts, as the AI system consistently applies the same detection criteria to all images regardless of how many were previously labeled, thereby improving both productivity and reliability
Solution Approach 2:
The AI object detection system performs self-service by automatically detecting and categorizing objects in images without requiring human intervention for category assignment. This self-service capability ensures consistent and unbiased search results across all queries, as the system independently applies learned object recognition patterns rather than relying on pre-assigned category codes that may be unevenly distributed
3Ease of operation
If word-based matching is used in text-based image search, then search engine can process queries simply, but irrelevant images are retrieved when text is related to query but not semantically meaningful
Solution Approach 1:
The patent replaces simple word-based string matching with AI-based semantic understanding. The system uses natural language processing to understand the semantic meaning of queries and object detection to identify objects in images, enabling the system to distinguish between superficial text matches and truly relevant semantic connections, thereby improving search result relevance while maintaining ease of operation
Solution Approach 2:
The patent changes the fundamental parameter of matching from textual string similarity to semantic object category similarity. By transforming both queries and image content into standardized object category representations through AI processing, the system achieves more accurate semantic matching that goes beyond simple word overlap, improving relevance while keeping the user interface simple
4Reliability
If ranking algorithms are applied after matching, then search engine can optimize results, but errors in matching operations limit final search quality even with excellent ranking algorithms
Solution Approach 1:
The patent applies preliminary action by performing AI object detection and category matching before the ranking stage. This preliminary filtering ensures that only images with semantically relevant objects are considered for ranking, thereby improving overall search quality. The system performs this preliminary action automatically using deep learning models, adding minimal complexity while significantly improving reliability
Solution Approach 2:
The patent substitutes the traditional two-stage process (matching then ranking) with an integrated AI-based approach where object detection and semantic matching occur simultaneously with ranking considerations. This substitution reduces the complexity of sequential processing by combining multiple functions into a unified AI-driven system, thereby improving search quality without proportionally increasing complexity
Data Source
AI summary
Provided is a method of image searching based on artificial intelligence (AI), the method including acquiring retrieved information, which includes at least one of a retrieved image and an image address, and a user query on the basis of a search result of an image search engine, detecting a keyword-category combination on the basis of a type of the acquired user query, determining whether cache data that matches the detected keyword-category combination exists, generating, in response to absence of the cache data that matches the keyword-category combination, an object-category combination through an AI technology based object detection on the acquired retrieved information.