Deformed Image Generation for Clearer Anomaly Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods face challenges in forming a balanced decision boundary for distinguishing between normal and defective data due to variations in the amount and quality of non-defective data, leading to ambiguity in the boundary between normal and defective samples.
Innovation Solution
A method for generating deformed images by augmenting original non-defective images, using neural networks to extract embedding vectors, calculate variance, and generate deformed images with added noise, while selecting optimal prompts and storing them in a non-defective data pool based on similarity measures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a decision boundary is formed by reflecting only the general characteristics of non-defects, then the decision boundary becomes too narrow, but if the diversity of non-defects is excessively reflected, the decision boundary becomes too wide and ambiguous
Solution Approach 1:
The patent applies dynamics by transforming static non-defective images into dynamic deformed images through controlled geometric transformations. The system generates multiple versions of normal images with various deformations (scaling, rotating, flipping, cropping) to create a more adaptable decision boundary that can handle diverse defect patterns while maintaining clarity.
Solution Approach 2:
The patent employs parameter changes by systematically varying image parameters such as scale factors, rotation angles, flip directions, and crop regions. These parameter transformations generate diverse training samples that expand the decision boundary's coverage of non-defect variations without sacrificing boundary precision, directly resolving the contradiction between adaptability and measurement precision.
2Reliability
If the amount and quality of non-defective data is increased to improve anomaly detection efficiency, then the decision boundary becomes more accurate, but the complexity of data processing and model training increases
Solution Approach 1:
The patent applies preliminary action by pre-processing normal images through augmentation before they are needed for training. The system proactively generates deformed versions of normal images in advance, creating an expanded training dataset that improves anomaly detection reliability without requiring complex real-time data processing during model training or inference.
Solution Approach 2:
The patent employs copying by creating multiple transformed copies of existing normal images through geometric transformations. Instead of requiring additional real-world non-defective data, the system generates synthetic copies with various deformations, effectively increasing data quantity and quality while avoiding the complexity of collecting and processing additional real data.
3Adaptability or versatility
If geometric transformations are applied to generate diverse training data, then the coverage of non-defect variations is improved, but the risk of generating unrealistic or overly distorted images increases
Solution Approach 1:
The patent applies local quality by selectively applying different types and intensities of geometric transformations to different regions or aspects of the training data generation process. The system uses localized transformation strategies that maintain realism in critical areas while introducing sufficient variation in other areas, balancing diversity with image quality and realism.
Data Source
AI summary
A method for generating deformed images includes receiving an original image including a non-defective object and augmenting the original image into a plurality of images, receiving the plurality of augmented original images and generating a deformed image from each of the augmented original images, receiving a name of an object included in the original image as text and generating an optimal prompt based on the name of the object, and determining whether to store the deformed image in a non-defective data pool based on the optimal prompt and the plurality of generated deformed images.