Image Mask Contour Augmentation for Small Labeled Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of maintaining data integrity in machine learning training with a small quantity of labeled samples, particularly due to high labeling costs, is addressed by the existing technologies.
Innovation Solution
An image augmentation device and method using a generative adversarial network (GAN) model to generate a large quantity of labeled sample images from a small quantity of labeled image masks through contour superimposition and augmentation parameters, including scaling, rotation, and movement, enhancing the training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large quantity of labeled samples are collected manually, then the integrity of training data is improved, but the labeling cost and time consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-processing images through contour extraction and mask generation before the main training process. The system extracts object contours from images, generates corresponding masks, and creates augmented samples in advance, so that when training data is needed, pre-processed and augmented samples are already available, significantly reducing the time required for data preparation and labeling during the actual training phase
Solution Approach 2:
The patent applies copying by using contour extraction to create mask representations of objects, and then using these masks to generate multiple augmented copies of training samples. The contour-based mask copying allows the system to create numerous variations of training data from a small set of original labeled images, maintaining data integrity while avoiding manual relabeling of each copy
2Quantity of substance
If manual labeling is performed extensively, then the quantity of labeled samples increases, but the labeling cost increases proportionally
Solution Approach 1:
The patent uses contour-based mask copying to generate multiple training samples from a single labeled image. By extracting the contour of labeled objects and creating mask representations, the system can produce numerous augmented samples without requiring manual labeling of each one, thereby increasing the quantity of labeled samples while keeping labeling costs constant
Solution Approach 2:
The patent applies parameter changes by transforming the original images through various parameter modifications including contour scaling, rotation, translation, and flipping. These parameter changes create diverse augmented samples from the original labeled images, increasing the quantity of training data without requiring additional manual labeling efforts
3Quantity of substance
If contour superimposition with multiple augmentation parameters is applied, then the quantity of generated samples increases, but the device complexity increases
Solution Approach 1:
The patent applies segmentation by breaking down the complex augmentation process into distinct modular steps: contour extraction, mask generation, parameter application (scaling, rotation, translation), and sample synthesis. This segmentation allows each component to be processed independently and efficiently, managing the complexity of generating multiple augmented samples through a structured, step-by-step approach rather than a monolithic complex process
Data Source
AI summary
An image augmentation device is provided, which includes a memory and a processor. The processor is configured for performing following operations: extracting a first object contour from a first image mask, wherein the first object contour corresponds to a first label; superimposing the first object contour to a superimposed region in a second image mask according to an augmentation parameter to generate a third image mask, wherein the augmentation parameter includes a contour scaling parameter, a contour moving distance, a contour rotation angle and a range which can be superimposed, and the third image mask includes the first object contour and a second object contour in the second image mask; and generating a sample image which corresponds to according to the first object contour and the second object contour in the third image mask by a generative adversarial network model for performing machine learning.