Bounding Box Correction for Geometric Image Transformation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Augmentation of input images for machine learning using geometric transformations, such as affine transformation, deforms bounding boxes, making it difficult to generate a good learned model due to the deformation of bounding boxes, which affects the accuracy of object detection models.

Innovation Solution

A learning device that includes an image transformation unit for performing geometric transformations, a bounding box correction unit that corrects the deformed bounding boxes by generating a new bounding box with a reduced area based on edge detection, and a machine learning unit that uses the corrected bounding boxes as training data to generate an object detection model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If geometric image transformation is performed for data augmentation, then the data count increases and data deviation is reduced, but the bounding box becomes deformed which deteriorates the learning model quality

Engineering Contradiction:
Improvedata countVSAvoidbounding box accuracy
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent applies preliminary action by performing bounding box correction before using the transformed image as training data. The correction unit predicts the original bounding box position based on the deformed bounding box and transformation parameters, restoring it to its pre-transformation state. This preliminary restoration ensures that the training data maintains accurate object location information while still benefiting from the geometric transformation for data augmentation.

Inventive Principle:
Principle #10Preliminary action

2Quantity of substance

If geometric image transformation is performed for data augmentation, then the data count increases and data deviation is reduced, but the learning model quality deteriorates due to bounding box deformation

Engineering Contradiction:
Improvedata countVSAvoidlearning model quality
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent applies preliminary action by performing bounding box correction before using the transformed image as training data. The correction unit predicts the original bounding box position based on the deformed bounding box and transformation parameters, restoring it to its pre-transformation state. This preliminary restoration ensures that the training data maintains accurate object location information while still benefiting from the geometric transformation for data augmentation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses transformation parameters as an intermediary to connect the deformed bounding box with the original bounding box. The correction unit utilizes these parameters (rotation angle, scaling factor, translation vector) to reverse the geometric transformation and recover the original bounding box coordinates, thereby maintaining reliability while allowing data transformation.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If deformed bounding boxes are used as training data, then the data count increases, but it becomes difficult to generate a good learned model

Engineering Contradiction:
Improvedata processing efficiencyVSAvoidobject detection accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent uses transformation parameters as an intermediary to connect the deformed bounding box with the original bounding box. The correction unit utilizes these parameters (rotation angle, scaling factor, translation vector) to reverse the geometric transformation and recover the original bounding box coordinates, thereby maintaining reliability while allowing data transformation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies preliminary action by performing bounding box correction before using the transformed image as training data. The correction unit predicts the original bounding box position based on the deformed bounding box and transformation parameters, restoring it to its pre-transformation state. This preliminary restoration ensures that the training data maintains accurate object location information while still benefiting from the geometric transformation for data augmentation.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4099264B1Learning device and learning method
Publication Date: 2024.04.24 SONY GROUP CORP
  • EP4099264B1 patent drawingFigure 1
  • EP4099264B1 patent drawingFigure 2
  • EP4099264B1 patent drawingFigure 3

AI summary

In a learning device, a transformation unit performs geometric image transformation on an input image including a hand image that is an image of a hand of a worker who is performing a product manufacturing work and a first bounding box set for an object to generate a transformed image that is the input image after the geometric image transformation, a correction unit sets a third bounding box obtained by correcting a second bounding box included in the transformed image as the transformed image, and a learning unit performs machine learning using the transformed image in which the third bounding box is set as training data to generate, with respect to a determination target image including the hand image, a learned model that outputs a plurality of objects defining each of a plurality of work states forming one process of the manufacturing work.