OOD Anomaly Detection With Context-Blended Image Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods struggle in real-world scenarios due to high variability in image contexts, leading to biased distributions and difficulties in specifying non-defective objects, particularly in manufacturing where traditional algorithms require homogeneous backgrounds and golden images, which are impractical and fail to handle diverse layouts and lighting conditions.
Innovation Solution
An anomaly detection method based on out-of-distribution (OOD) using context augmentation and multi-geometry projection networks, where objects are segmented from their contexts, similarities are calculated, and blended images are generated to train a detection model, enabling robust anomaly detection across varied backgrounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional anomaly detection methods use golden images for comparison, then detection accuracy can be maintained in controlled settings, but the method fails when image contexts have high variability such as different layouts, lighting conditions, and camera settings
Solution Approach 1:
The patent segments the image into object region and background region separately. The object region contains the component to be inspected, while the background region contains contextual information. This segmentation allows the system to handle varied backgrounds by treating them as separate from the object, enabling anomaly detection to work across different layouts, lighting conditions, and camera settings without requiring matched golden images for each context variation.
Solution Approach 2:
The patent extracts the object region from the background region using segmentation. By extracting only the relevant object portion and separating it from the variable background, the system can perform anomaly detection on the object independent of context variations. This extraction eliminates the need for golden images that must match the exact background context, as the background is removed and treated separately.
2Measurement precision
If anomaly detection requires homogeneous backgrounds and golden images, then detection can be performed in constrained settings, but it becomes impractical for real-world manufacturing with diverse layouts and lighting conditions
Solution Approach 1:
The patent creates a universal anomaly detection system that works across multiple contexts and conditions. By segmenting objects from backgrounds and using region-swapping augmentation to train on diverse background variations, the single detection model becomes universally applicable to different layouts, lighting conditions, and camera settings without requiring separate golden images for each scenario, making it feasible for real-world manufacturing environments.
Solution Approach 2:
The patent performs preliminary data augmentation during training by swapping object regions across images with different backgrounds. This preliminary action of pre-training on augmented data with varied backgrounds prepares the model to handle real-world context variations, eliminating the need for homogeneous backgrounds during actual deployment and making the system practical for diverse manufacturing settings.
3Productivity
If the model is trained on data with limited background variability, then training is simpler and faster, but the model's robustness degrades significantly when encountering highly varied contexts
Solution Approach 1:
The patent performs preliminary augmentation during training by swapping object regions across images with different backgrounds to create diverse training samples. This preliminary action exposes the model to varied contexts during training, improving robustness to background variations in deployment without requiring complete retraining for each new context, thus maintaining training efficiency while enhancing reliability.
Solution Approach 2:
The patent changes the background parameter during training by swapping backgrounds across different images through region swapping augmentation. This parameter change exposes the model to diverse background conditions during training, improving its robustness to background variations in real-world deployment while maintaining the same object detection task, thus enhancing reliability without fundamentally changing the training process.
Data Source
AI summary
This anomaly detection method, based on out-of-distribution techniques, is executed by a computing device. It starts by obtaining a training dataset containing various images, including a first image and multiple second images. The method segments objects and contexts in each image, calculating the similarity between the object in the first image and those in the second images. A candidate image is selected if its similarity exceeds a predefined threshold. The object from the first image is blended with the context of the candidate image to produce a blended image. A detection model is then trained using this dataset. Subsequently, in-distribution embeddings are generated, and a test embedding is created. The test sample is classified as an anomaly when the minimum distance between the in-distribution embeddings and the test embedding exceeds a default value.