Recognition of Chevrons Using AI
The Task
A chevron manufacturer faced a significant challenge in their marketplace operations. The company had an extensive product catalog—thousands of different chevrons with unique characteristics such as color, size, shape, text, and symbolism.
When listing products on marketplaces, it was crucial to accurately match product photographs with corresponding SKUs in the catalog.
Manual matching required significant time and effort while being prone to human error. Traditional computer vision systems demanded substantial investments in development and model training, which did not align with the company's budgetary constraints.
When listing products on marketplaces, it was crucial to accurately match product photographs with corresponding SKUs in the catalog.
Manual matching required significant time and effort while being prone to human error. Traditional computer vision systems demanded substantial investments in development and model training, which did not align with the company's budgetary constraints.
Technical solution
System Architecture
A cutting-edge system based on large language models (LLMs) was developed, which processes product images and generates:
- Structured parametric descriptions - detailed textual characteristics of each chevron
- Vector embeddings - mathematical representations for efficient similarity-based search
Technology Stack
- Backend: FastAPI (Python)
- Database: PostgreSQL with the pgvector extension
- LLM: OpenRouter.io API for image processing
- Vectorization: SentenceTransformer for generating embeddings
- Frontend: HTML/JavaScript with Jinja2 templates
How It Works
Key Components
LLMService: Processes images through a Large Language Model API, generating detailed product descriptions in a structured format.
VectorService: Converts textual descriptions into vector embeddings for mathematical similarity search.
DBService: Manages data storage in PostgreSQL with support for vector operations via pgvector.
ImportModule: Automates the process of importing data from Excel files and processing images.
Workflow
- Data Import: The system loads product information from Excel files along with corresponding images.
- Image Processing: The LLM analyzes each photo and generates a detailed description of the chevron's characteristics.
- Vectorization: The textual descriptions are converted into vector embeddings.
- Indexing: Data is stored in the database with support for fast vector similarity search.
- Search: When a new image is uploaded, the system finds the most similar products.
Testing Results
The system was tested on the company's real-world data:
- Top-1 Accuracy: 77% - the system correctly identifies the exact match in 77% of cases
- Overall Accuracy: 95.2% - the correct product is found within the top-5 results in 95.2% of cases
- Processing Speed: Less than 2 seconds per query
- Scalability: The system successfully handles catalogs containing thousands of items
Business Impact
- Time Savings
- Reduction of matching time from 5-10 minutes to 30 seconds per product
- Automation of 95% of matching processes
- Accuracy Improvement
- 80% reduction in matching errors
- Enhanced data quality on marketplaces
- Cost Reduction
- Leveraging pre-trained LLMs instead of developing custom models
- Minimal computational resource requirements
Implementation Features
- Configurability
- Customizable search parameters and similarity thresholds
- Support for various input data formats
- Capability for fine-tuning on client-specific data
- Monitoring and Analytics
- Detailed logging of all search operations
- Accuracy and performance statistics
- Tools for results analysis
- Scalability
- Asynchronous request processing
- Support for batch data processing
- Horizontal component scaling
Key Advantages
- Cost-Effectiveness: Utilizing pre-trained LLMs significantly reduces development costs compared to building proprietary computer vision models.
- Rapid Deployment: The system can be deployed and configured within days.
- High Accuracy: The combination of LLM and vector search delivers precision comparable to expensive specialized solutions.
- Versatility: The approach can be adapted to other product types with distinct visual characteristics.
Development Roadmap
- Integration with Product Information Management (PIM) systems
- Expansion to additional product categories
- Implementation of automated categorization features
- Integration with marketplace APIs for automated product listing
We'll find the right solution for your business