Synthesis Unit for Retail Product Image Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting product shortage or display disturbance in stores face challenges due to misidentification caused by varying showcase configurations and product orientations, leading to deteriorated detection accuracy and the difficulty in capturing high-quality training data for each store.
Innovation Solution
A training data generation system that includes a shelf-image acquisition unit, a product-image acquisition unit, and a synthesis unit, which generates training data by synthesizing shelf images and product images, adjusting the product display based on the shape of the shelf or product to match the specific store environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a training model is trained using images captured at one specific location, then the training process is simple and efficient, but misidentification occurs in different stores and detection accuracy deteriorates
Solution Approach 1:
The patent creates virtual copies of product images and synthesizes them with background images from different store locations. Instead of capturing real images from every store, the system generates synthetic training images that replicate various store environments, allowing the model to learn from diverse scenarios without physical presence in each location.
Solution Approach 2:
The synthesis unit varies parameters such as product orientation, display position, and background characteristics to generate diverse training images. By changing these parameters systematically, the model learns to recognize products under various conditions while maintaining training efficiency through automated image generation.
2Measurement precision
If high-quality training data is captured for each store individually, then detection accuracy improves, but the complexity and time required for data collection increases significantly
Solution Approach 1:
The system uses a universal product image database that can be applied across multiple stores. Instead of creating separate training datasets for each store, the same product images are synthesized with different store backgrounds, making the training data collection process universal and applicable to any store location.
Solution Approach 2:
The synthesis unit acts as an intermediary between product images and store backgrounds. It combines these elements to create realistic training images without requiring direct capture in each store environment, simplifying the data collection process while maintaining accuracy.
3Ease of operation
If product images are captured with fixed display orientations, then the capture process is simple, but the model cannot adapt to different product orientations in various stores
Solution Approach 1:
The synthesis unit dynamically rotates and repositions product images to create training samples with various orientations. This dynamic transformation allows the model to learn product recognition across different orientations while the actual image capture process remains simple and fixed.
Data Source
AI summary
A training data generation device comprises a shelf-image acquisition unit, a product-image acquisition unit, and a synthesis unit. The shelf-image acquisition unit acquires a shelf image constituting one compartment of a shelf on which a product is displayed. The product-image acquisition unit acquires a product image of the product displayed on the shelf. The synthesis unit generates training data by synthesizing the shelf image and the product image, and the synthesis unit additionally, in accordance with the shape of the shelf and/or the shape of the product, causes the display in the product image to differ and synthesizes the result with the shelf image.