AI Training Dataset Expansion via Image Rotation and Flipping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI model training datasets are often limited by equipment and cost constraints, resulting in insufficient pictures for training, which reduces the accuracy of AI deep learning models.
Innovation Solution
A device and method that derive additional pictures from an original picture by flipping and rotating it, along with annotation processes, to increase the dataset size for AI deep learning model training, using a processor and storage system to generate amplification pictures and store them with unified annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If AI model training uses limited equipment and cost constraints, then equipment cost and time are reduced, but the number of training pictures is insufficient, reducing model accuracy
Solution Approach 1:
The patent applies the copying principle by generating synthetic training pictures through image processing operations (flipping, rotating, cropping, resizing) on existing original pictures. This creates multiple copies of training data from a single source picture, significantly increasing the quantity of training pictures without requiring additional physical equipment or increasing production line complexity.
Solution Approach 2:
The patent applies parameter changes by modifying image parameters such as orientation (flipping horizontally/vertically), position (cropping different regions), and dimensions (resizing). These parameter transformations generate varied training pictures from the same original image, effectively increasing dataset diversity and quantity without additional hardware costs.
2Quantity of substance
If more original pictures are collected from production line, then model accuracy improves, but cost and time increase
Solution Approach 1:
The patent applies preliminary action by pre-processing original pictures into multiple derived pictures before model training. The image processing operations (flipping, rotating, cropping, resizing) are performed in advance to expand the training dataset, eliminating the need to wait for additional original pictures to be collected from the production line during or after training.
Solution Approach 2:
By creating synthetic copies of training pictures through image transformations, the patent rapidly expands the training dataset without the time-consuming process of collecting additional original pictures from the production line. Multiple training pictures are generated from each original picture through copying and transformation operations.
3Quantity of substance
If image processing operations are applied to original pictures, then number of training pictures increases, but data processing complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into separate, simple operations: flipping (horizontal/vertical), rotating (90-degree increments), cropping (region selection), and resizing. These segmented operations are computationally simple and can be applied independently to original pictures to generate derived pictures, avoiding complex processing while expanding the dataset.
Solution Approach 2:
The patent uses simple parameter changes (orientation, position, dimensions) that can be implemented through basic image processing functions. These parameter transformations increase training picture quantity without requiring complex processing algorithms, maintaining low data processing complexity while effectively expanding the training dataset.
Data Source
AI summary
A method for deriving further and additional pictures from an original picture, for Artificial Intelligence (AI) training purposes, is applied in a device. The device establishes an original picture set and sets the original pictures as a training picture set for AI training. The original pictures are rotated or flipped or both to obtain amplification pictures. The original pictures are annotated, and each of the amplification pictures is annotated according to a preset conversion rule. The original pictures, the amplification pictures, the annotated original pictures, and the annotated amplification pictures are stored, for inclusion in the AI training picture set.