Visual Anomaly Detection Using Self-Supervised Nominal Part Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based computer vision models for defect detection in manufacturing are limited by the need for labor-intensive manual labeling of rare defective data, subjectivity in labeling, and inability to detect unknown defects or defects in varying regions, especially in industrial environments with high within-class variance.
Innovation Solution
A self-supervised anomaly detection framework that utilizes a pre-trained loss computation neural network to learn feature representations from real-world conditions, combined with a main anomaly detection neural network trained on nominal data to reconstruct images and measure perceptual loss, enabling robust detection of defects without requiring labeled defective data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with manual labeling is used to train deep learning models for defect detection, then the model can accurately detect known defects in similar regions, but it cannot detect unknown defects or defects in different regions and requires extremely long data collection periods and labor-intensive manual labeling
Solution Approach 1:
The patent inverts the traditional supervised learning approach by using unsupervised learning where the model is trained only on normal defect-free images. The model learns what normal appearance looks like and automatically detects anomalies without any manual labeling of defective images, thereby eliminating the time-consuming data collection and labeling process while maintaining detection capability
Solution Approach 2:
The system performs self-supervised learning where the model automatically generates its own training signals from normal images through data augmentation and self-supervised pre-training. The anomaly detection network learns to reconstruct normal patterns and automatically identifies deviations as defects, making the system self-sufficient without external labeling resources
2Adaptability or versatility
If supervised learning methods are used to train models on limited defective data, then the model can capture known defects, but it cannot accurately detect unknown types of defects or known defects showing up in different regions due to high within-class variance
Solution Approach 1:
The patent applies self-supervised pre-training on normal images before the actual anomaly detection task. This preliminary action allows the model to learn robust feature representations of normal variations and patterns, enabling it to adapt to different defect types and regions while maintaining high detection accuracy through the contrast between learned normal patterns and actual anomalies
Solution Approach 2:
The anomaly detection model becomes universal by learning the underlying normal distribution of the specific manufacturing process through self-supervised learning. This single model can detect multiple types of defects across different regions without retraining, as it identifies any deviation from the learned normal pattern regardless of defect type or location
Data Source
Figure 1
Figure 2(a)~2(c)
Figure 3
AI summary
An AI-based method for visual inspection of parts manufactured on a shop floor includes acquiring a set of real images of nominal parts manufactured on the shop floor to create training datasets. A self-supervised pre-trainer module is used to pre-train a loss computation neural network in a self-supervised learning process using a first dataset on pretexts defined by real-world conditions pertaining to the shop floor. The first dataset is labeled by automatically extracting pretext-related information from image metadata. A main anomaly trainer module is used to train a main anomaly detection neural network to reconstruct a nominal part image from an input manufactured part image in an unsupervised learning process using a second dataset. The main anomaly training measures a perceptual loss between an input image and a reconstructed image by measuring a difference between feature representations thereof at one or more layers of the pre-trained loss computation neural network.