Synthesizing Training Data for Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches to training machine learning algorithms, such as convolutional neural networks, require a large amount of high-quality training data, which is expensive and time-consuming to obtain, especially for classifying content in electronic environments where granular classes and sub-classes are involved, leading to inefficiencies and inaccuracies due to the need for extensive data labeling and handling of background artifacts.
Innovation Solution
The method involves synthesizing training images by removing background pixels from catalog images and blending item representations into various backgrounds to generate diverse training data, using morphological operations to reduce artifacts, and combining real-world and synthesized images to ensure neural network accuracy across different classes and sub-classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional approaches are used to obtain training data through extensive data labeling, then classification accuracy can be improved, but the time and resource requirements increase significantly
Solution Approach 1:
The patent creates synthesized training images by copying and pasting item representations from catalog images into various background images. This copying approach generates diverse training data without requiring manual labeling of new images, thus maintaining classification accuracy while significantly reducing data labeling time and resources
Solution Approach 2:
The patent performs preliminary processing of catalog images to extract item representations and remove backgrounds before synthesis. By preparing these components in advance, the system enables rapid generation of training images without requiring time-consuming manual labeling during the training data collection phase
2Ease of manufacture
If catalog images with solid backgrounds are used directly for training, then data acquisition is simplified, but background artifacts negatively impact neural network accuracy
Solution Approach 1:
The patent extracts item representations from catalog images by removing the solid backgrounds through image processing techniques. This extraction separates the item content from the problematic background, allowing the items to be synthesized into diverse backgrounds that do not introduce artifacts, thereby maintaining both acquisition simplicity and neural network accuracy
Solution Approach 2:
The patent introduces synthesized background images as an intermediary element between the extracted item representations and the final training images. These intermediary backgrounds provide variety and realism without introducing the artifacts present in original catalog backgrounds, thus resolving the conflict between simplicity and accuracy
3Adaptability or versatility
If a large variety of classes and sub-classes are targeted, then comprehensive content classification is achieved, but the data collection and classification burden becomes daunting
Solution Approach 1:
The patent creates a universal synthesis framework that can generate training images for multiple classes and sub-classes using the same item extraction and background composition techniques. This multi-functional approach allows comprehensive classification coverage without proportionally increasing data classification complexity, as the synthesis process handles diverse classes uniformly
Solution Approach 2:
The patent applies the copying and synthesis technique across multiple classes and sub-classes, replicating the successful approach for each category. This systematic copying strategy enables comprehensive content classification while maintaining manageable complexity through automation of the synthesis process
Data Source
AI summary
Training images can be synthesized in order to obtain enough data to train a convolutional neural network to recognize various classes of a type of item. Images can be synthesized by blending images of items labeled using those classes into selected background images. Catalog images can represent items against a solid background, which can be identified using connected components or other such approaches. Removing the background using such approaches can result in edge artifacts proximate the item region. To improve the results, one or more operations are performed, such as a morphological erosion operation followed by an opening operation. The isolated item portion then can be blended into a randomly selected background region in order to generate a synthesized training image. The training images can be used with real world images to train the neural network.