Indicia Scanner Tile Selection for ML Training Data Harvesting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The lack of sufficient high-quality training data and the inefficiency of data transmission due to limited memory and processing capabilities in indicia scanners hinder the development of robust machine learning models for indicia scanning applications.

Innovation Solution

An indicia scanner captures images, segments them into tiles, determines quality scores, and transmits selected image tiles to a central server for training data generation, where clean images are generated and used to train the ML model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of energy

If image segmentation and quality scoring is performed at the indicia scanner, then bandwidth requirements are reduced, but processing complexity at the scanner increases

Engineering Contradiction:
Improvebandwidth requirementsVSAvoidprocessing complexity at scanner
Core Design Contradiction:
Loss of energyVSDevice complexity

Solution Approach 1:

The patent divides the image into multiple tiles and performs quality scoring on individual tiles rather than processing the entire image. This segmentation allows the scanner to identify and transmit only relevant high-quality tiles to the server, reducing bandwidth consumption while distributing processing tasks efficiently between the scanner and server.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary quality assessment and segmentation at the scanner before transmission. By pre-processing images to identify high-quality tiles locally, the scanner reduces the amount of data that needs to be transmitted, thereby reducing bandwidth requirements while the added processing complexity is offset by the efficiency gains in data transmission.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If the ML model is optimized for a specific domain, then prediction accuracy improves, but the model requires more specialized training data

Engineering Contradiction:
Improveprediction accuracyVSAvoidtraining data requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent focuses on collecting and transmitting only high-quality tiles that are relevant to specific indicia types and scanning conditions. By curating training data with local quality (focusing on specific domains like postal indicia, retail barcodes, etc.), the system achieves better prediction accuracy for those specific domains without needing excessively large datasets covering all possible scenarios.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system enables the indicia scanner to autonomously collect, segment, quality-score, and transmit relevant training data to the server. This self-service approach allows the system to automatically generate domain-specific training data without manual intervention, reducing the burden of data collection while improving model accuracy for specific scanning domains.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3989180B1Methods and systems of harvesting data for training machine learning (ML) model
Publication Date: 2025.11.26 HAND HELD PRODS INC
  • EP3989180B1 patent drawingFigure 1
  • EP3989180B1 patent drawingFigure 2
  • EP3989180B1 patent drawingFigure 3

AI summary

Various embodiments disclosed herein describe a method comprising receiving indicia data from an indicia scanner. The indicia data comprises at least decoded data obtained based on decoding an indicium in an image, an image tile comprising a portion of the indicium, and/or location of one or more corners of the portion of the indicium in the image. Further, the method includes generating an image of an ideal indicium based on at least the decoded data. Thereafter, the image of the ideal indicium is modified to generate a modified image of the ideal indicium. Further, the portion of the indicium is retrieved from modified image. A clean image tile comprises a portion of the ideal indicium. Furthermore, the method includes generating training data, wherein the training data includes the portion of the indicium and the portion of the ideal indicium.