Automated Foreign-Material Sample Generation from Single-Class Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems for image classification, particularly in mono-substance processing streams, face challenges in training models when samples from both classes are not readily available, leading to inefficiencies and high costs in obtaining pre-classified training images, especially for the 'foreign material' class.
Innovation Solution
An Artificially Spiked Data Generation System (ASDGS) uses a single class of data to generate artificially spiked data for the second class by combining a dataset of clean images with object, shape, and texture libraries, applying calibration operations to create realistic and diverse training images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If manual generation of class B samples using Photoshop or GIMP is used, then some training images for the unavailable class can be created, but the process is time-consuming, labor-intensive, and produces insufficient non-natural images that are ineffective for training
Solution Approach 1:
The system creates synthetic training images by copying and combining real images from the available class (class A) with foreign object images to generate realistic class B samples. This automated copying process replaces manual Photoshop/GIMP operations, producing sufficient training images without time-consuming manual labor while maintaining image naturalness through real image sources
Solution Approach 2:
The patent replaces the mechanical manual process of using Photoshop or GIMP with an automated computerized image synthesis system. The mechanical system involves manual manipulation of images by operators, while the invention substitutes this with automated algorithms that programmatically combine images, eliminate human labor, and scale efficiently to generate large numbers of training samples
2Reliability
If pre-classified training images from both classes are obtained through traditional methods, then the classification model can be trained effectively, but it requires significant cost and effort to obtain samples from both classes
Solution Approach 1:
The system copies images from the readily available class A dataset and combines them with foreign object images to synthesize class B training samples. This copying approach eliminates the need to manually collect, annotate, and classify real class B images, significantly reducing cost and effort while providing sufficient training data for effective model training
Solution Approach 2:
The patent uses an automated image synthesis system as an intermediary between the available class A images and the needed class B training samples. This intermediary process automatically generates the missing class B samples through computational combination, avoiding the need for direct manual collection and classification of expensive-to-obtain foreign material images
3Ease of operation
If a confidence rating system is used to determine class membership, then some classification capability is achieved, but it still requires hand-labeling of the dataset which is labor-intensive
Solution Approach 1:
The system performs preliminary action by pre-generating synthetic class B training images before model training begins. This advance preparation eliminates the need for time-consuming hand-labeling during the training process, as the training dataset is already prepared with accurate class labels embedded in the synthetic image generation process
Solution Approach 2:
The patent uses automated copying and combination of images to create training datasets with known labels, replacing the hand-labeling process. The synthetic image generation inherently provides ground truth labels based on the combination process, eliminating manual annotation time while maintaining classification capability
Data Source
AI summary
An ASDGS (Artificially Spiked Data Generation System) may comprise computing and mechanical systems for using a single class of data to generate artificially “spiked” data of a second class. To generate each spiked image, the ASDGS may randomly select a clean image an augmentation object (“AO”) from an object library, shape library, and/or hair library. The ASDGS may use a texture library to add or change the texture of the AO. The ASDGS may adjust the lighting and coloring of the AO to be similar to the clean image, and may then add the AO to the clean image to generate a spiked image.