Synthetic Background Generation for Object Detection Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing object detection algorithms require high-quality supervised data that includes background images, which may not be available in actual use scenarios, hindering the effectiveness of target detection.

Innovation Solution

A supervised data generation program that selects a target image, performs a transformation process to generate a background image, and combines the transformed target image with the background image to create supervised data, enabling training of object detection models without actual background images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If actual background images are used for training object detection models, then the detection accuracy is improved, but the applicability is reduced because actual background images may not be available in all use scenarios

Engineering Contradiction:

Improvedetection accuracyVSAvoidapplicability

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent creates synthetic background images by copying and transforming target images from the training dataset. Instead of requiring actual background images, the system generates artificial backgrounds by applying geometric transformations (rotation, scaling, flipping) and composite operations to target images, then uses these synthesized backgrounds for training the object detection model, achieving both detection accuracy and broad applicability

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent makes the target images themselves serve as the source material for creating backgrounds. The training data automatically generates its own background images through transformation processes, eliminating the need for external background image datasets and enabling the system to be self-sufficient in data preparation

Inventive Principle:

Principle #25Self-service

2Productivity

If transformation processes are applied to generate background images, then the productivity of data generation is improved, but the manufacturing precision may deteriorate due to potential loss of image quality

Engineering Contradiction:

Improvedata generation efficiencyVSAvoidimage quality

Core Design Contradiction:

ProductivityVSManufacturing precision

Solution Approach 1:

The patent applies transformation processes selectively and with controlled intensity. Not all target images undergo all possible transformations, and the system maintains original images alongside transformed versions. This partial application of transformations balances data augmentation benefits with preservation of original image quality, achieving both high productivity and acceptable manufacturing precision

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20250045876A1Non-transitory storage medium storing supervised data generation program, supervised data generation method, supervised data generation apparatus, training apparatus, and data structure of supervised data

Publication Date: 2025.02.06 MICWARE CO LTD

AI summary

A technique generates supervised data without a background image in actual use. Supervised data generation program is a program for generating supervised data to generate a trained model for outputting a result from identifying a target in response to input image data of an image including the target corresponding to a target image. The program causes a computer to perform operations including selecting a first target image from an image group including a plurality of different target images and performing a transformation process to generate a background image, and selecting a second target image from the image group and combining the second target image with the background image to generate supervised data.