Training Data Generation for Overlapping Product Identification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image processing and machine learning systems face difficulties in distinguishing between overlapping products in images, which hinders accurate product identification and quantity computation.

Innovation Solution

A training data generation method that creates learning group images by randomly arranging individual images of products, allowing for partial overlap, and assigns labels with quantity, positional, or centroid information to enable the computing unit to identify and differentiate overlapping products.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional image processing or machine learning is used to identify products, then the system can process images efficiently, but it cannot distinguish between overlapping products

Engineering Contradiction:
Improveproduct identification accuracyVSAvoidhandling of overlapping products
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies preliminary action by pre-processing individual product images before combining them into group images. Individual images are captured, processed, and stored separately with their identification data, allowing the system to later assemble them in various overlapping configurations. This preparation enables the machine learning model to learn from pre-processed individual product characteristics even when they appear overlapped in group images.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If training data is generated using real overlapping product images, then the model learns realistic scenarios, but data collection becomes time-consuming and complex

Engineering Contradiction:
Improvetraining data realismVSAvoiddata collection time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent uses copying by creating synthetic group images from individual product images through digital assembly rather than capturing real overlapping scenes. Individual product images are copied and combined in various overlapping arrangements to generate training data. This approach maintains the realism of product appearance while eliminating the time-consuming process of manually arranging physical products for each training image.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system applies self-service by automatically generating training data from individually captured product images without requiring manual intervention for each group image creation. The machine automatically assembles individual images into overlapping group configurations, assigns appropriate labels, and prepares training datasets, eliminating the need for manual data collection and processing efforts.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If individual product images are captured and processed separately, then identification accuracy improves, but the system cannot handle multiple products in a group image

Engineering Contradiction:
Improveindividual product recognition accuracyVSAvoidgroup image processing capability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies segmentation by dividing the group image processing task into individual product image processing steps. Each product in a group image is first captured and processed as an individual image, extracting its identification characteristics separately. These individually processed images are then combined into group images for training, allowing the model to learn both individual product recognition and group context simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent merges individually processed product images into group images for comprehensive training. After individual images are processed and characterized, they are combined into group images that represent realistic shopping scenarios. This merging allows the machine learning model to learn from both individual product features and their relationships within groups, improving both accuracy and productivity.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3651067A1Training data generation method, training data generation apparatus, and product identification apparatus
Publication Date: 2020.05.13 ISHIDA CO LTD
  • EP3651067A1 patent drawingFigure 1
  • EP3651067A1 patent drawingFigure 2
  • EP3651067A1 patent drawingFigure 3

AI summary

[Problem] To enable overlapping products to be distinguished. [Solution] A method of generating training data 40 is a method of generating training data 40 used to generate a computing unit X for a product identification apparatus 10 that computes, from a group image in which there are one or more types of products G, the quantities of each of the products G included in the group image. The training data 40 includes plural learning group images 41 and labels 42 assigned to each of the plural learning group images 41. The method of generating the training data 40 includes a first step of acquiring individual images 43al to 43a6, 43b1 to 43b6, 43c1 to 43c6 in each of which there is one product G of one type and a second step of generating the plural learning group images 41 including one or more of the products G by randomly arranging the individual images. The plural learning group images 41 generated in the second step include learning group images 41 in which the individual images at least partially overlap each other.