Synthetic Training Data for Barcode-Free Product Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional contactless sales systems rely on barcodes or SKUs, which can lead to delays, errors, and inefficiencies in identifying and adding products to inventory due to the need for precise orientation and potential scanner failures, resulting in incorrect shipments and inventory inaccuracies.
Innovation Solution
A system that uses machine learning models trained with synthetic datasets generated by combining object assets and background images, allowing products to be recognized based on their plain appearance without barcodes, and optimizing the training process through techniques like class balancing to reduce bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If barcodes or SKUs are used for product identification, then product identification can be achieved, but delays and inaccuracies occur due to orientation requirements, physical wear, and potential misidentification
Solution Approach 1:
The patent extracts the identification function from traditional barcodes/SKUs and implements it through machine learning models that recognize products based on their visual appearance. This removes the dependency on orientation-sensitive, wear-prone barcode systems and replaces them with robust image-based recognition that can identify products from multiple angles and conditions.
Solution Approach 2:
The patent replaces the mechanical barcode scanning system with an optical image processing system using cameras and machine learning algorithms. This substitution eliminates the need for precise orientation and physical contact required by traditional barcode scanners, thereby reducing identification delays and improving accuracy.
2Ease of operation
If traditional barcode scanning systems are used, then product identification is possible, but the system complexity increases due to orientation requirements and additional scanning equipment
Solution Approach 1:
The patent implements a universal image-based recognition system that can identify products regardless of their orientation, position, or condition. This multi-functional approach replaces multiple specialized barcode scanners with a single camera system that handles all identification tasks through machine learning, simplifying the overall system architecture.
Solution Approach 2:
Instead of requiring the product to be presented in a specific orientation to the scanner (traditional approach), the patent inverts the problem by training the system to recognize products in multiple orientations simultaneously. This allows the identification system to work with products in any position, greatly simplifying the operation.
3Reliability
If synthetic data generation is implemented, then machine learning model training is improved, but data processing time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by generating synthetic training data in advance through computer-generated images of products in various orientations, lighting conditions, and positions. This pre-generated synthetic data is then used to train the machine learning models before deployment, improving model reliability without requiring extensive processing time during actual product identification operations.
Data Source
AI summary
Systems and methods are provided for generating synthetic datasets. The system can generate a plurality of object assets, wherein each asset comprises an object of interest. A plurality of asset classes can be defined, wherein each class comprises a subset of the plurality of assets depicted in a target zone, and wherein each asset of the subset is depicted one or more times in the target zone. The system can determine, for each asset of the plurality of object assets, a representation of a number of times the asset is depicted in an asset class of the plurality of asset classes and determine one or more differences between the representations. At least one random class can be defined, wherein the at least one random class comprises one or more assets of the plurality of object assets to reduce the one or more differences between the representations.