Image Processing Device for Automated Training Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current supervised learning methods face challenges in increasing the number of training data and improving classification accuracy due to the high cost and time required for manually labeling the position and shape of objects in images, and the limitations of semi-supervised learning in generating new images with the same correct answer label.
Innovation Solution
An image processing device and method that automatically generates a new processing target image by assigning correct answer labels to learning images, segmenting them, and regenerating classifier data based on classification scores, allowing for increased training data without manual intervention and improving classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to provide training data with correct answer labels and object position/shape information, then classification accuracy can be improved, but the cost and time required increases significantly
Solution Approach 1:
The patent generates synthetic training data by copying and transforming existing labeled images through geometric transformations (rotation, scaling, flipping) and adding noise. This creates multiple variants of original images with automatically generated correct answer labels, eliminating the need for manual labeling of each variant while preserving the semantic content and labels of the source images.
Solution Approach 2:
The patent segments images into multiple regions or crops to create separate training samples from a single original image. By dividing an image into different segments and treating each as an independent training sample, the system increases the quantity of training data without requiring additional manual labeling effort.
2Measurement precision
If the number of training data is increased to improve discrimination accuracy, then classification performance improves, but the cost of manually assigning correct answer labels increases
Solution Approach 1:
The system creates multiple copies of existing labeled images by applying various transformations and noise additions. Each copied image serves as a new training sample with an automatically generated correct answer label derived from the original image's label, significantly increasing training data quantity without proportional increases in labeling cost.
Solution Approach 2:
The patent transforms existing training images by modifying parameters such as rotation angle, scaling factor, brightness, contrast, and noise levels. These parameter changes generate diverse image variants from a single source image, expanding the training dataset while the correct answer labels are automatically adjusted based on the transformation applied.
3Quantity of substance
If semi-supervised learning is used to generate new images from labeled images, then the number of training data can be increased, but it is impossible to generate a large number of new images without manual intervention
Solution Approach 1:
The system automatically generates numerous image copies through programmatic transformations including geometric operations (rotation, translation, scaling), filtering operations (blur, sharpen, noise addition), and color space transformations. This fully automated copying process creates large volumes of synthetic training data without human intervention.
Solution Approach 2:
The patent implements dynamic image generation where parameters such as transformation angles, scaling factors, and noise levels are varied randomly or systematically to create diverse image variants. This dynamic approach allows the system to generate an effectively infinite number of training samples from a limited set of original labeled images.
Data Source
AI summary
An image processing device includes an input reception section that receives a learning image and a correct answer label, a processing section that performs a process that generates classifier data and a processing target image, and a storage section. The processing section generates the processing target image that is the entirety or part of the learning image, calculates a feature quantity of the processing target image, generates the classifier data based on training data that is a set of the feature quantity and the correct answer label assigned to the learning image that corresponds to the feature quantity, generates an image group based on the learning image or the processing target image, classifies each image of the image group using the classifier data to calculate a classification score of each image, and regenerates the processing target image based on the classification score and the image group.