Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows (2008.12577v1)

Abstract: The detection of manufacturing errors is crucial in fabrication processes to ensure product quality and safety standards. Since many defects occur very rarely and their characteristics are mostly unknown a priori, their detection is still an open research question. To this end, we propose DifferNet: It leverages the descriptiveness of features extracted by convolutional neural networks to estimate their density using normalizing flows. Normalizing flows are well-suited to deal with low dimensional data distributions. However, they struggle with the high dimensionality of images. Therefore, we employ a multi-scale feature extractor which enables the normalizing flow to assign meaningful likelihoods to the images. Based on these likelihoods we develop a scoring function that indicates defects. Moreover, propagating the score back to the image enables pixel-wise localization. To achieve a high robustness and performance we exploit multiple transformations in training and evaluation. In contrast to most other methods, ours does not require a large number of training samples and performs well with as low as 16 images. We demonstrate the superior performance over existing approaches on the challenging and newly proposed MVTec AD and Magnetic Tile Defects datasets.

• The paper proposes DifferNet, which leverages normalizing flows as a density estimator for CNN-extracted features to detect defects using only non-defective samples.
• It integrates a multi-scale feature extractor to preserve high-dimensional details, enabling precise anomaly localization across diverse defect types.
• Empirical evaluations on MVTec AD and MTD datasets demonstrate superior ROC performance compared to traditional anomaly detection methods.

Semi-Supervised Defect Detection with Normalizing Flows

The paper "Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows" introduces a novel approach for defect detection in manufacturing processes using a method called DifferNet. The authors, Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn from Leibniz University Hanover, propose a technique leveraging normalizing flows to address the challenges associated with detecting rare and a priori unknown defects within fabricated products. This approach is notable for its semi-supervised nature, relying exclusively on non-defective training samples for model development.

Methodology

The core of DifferNet is the use of normalizing flows as a density estimator for features extracted from convolutional neural networks (CNNs). By mapping input image features to a latent space via a bijective transformation, regularized by a Gaussian distribution, the method assigns likelihood scores to the image data. Images associated with anomalies exhibit low likelihoods, enabling defect detection through a derived anomaly score. This score integrates likelihood observations from various transformed versions of the input image, enhancing robustness against diverse anomaly types and input variations.

A key innovation is the incorporation of a multi-scale feature extractor that preserves high-dimensional feature richness while translating it into a form digestible by normalizing flows, typically suited for lower-dimensional data distributions. This multi-scale approach significantly contributes to the descriptive capabilities of the feature space, improving both anomaly detection and localization without requiring extensive datasets.

Empirical Evaluation

The DifferNet was evaluated on the MVTec AD and Magnetic Tile Defects (MTD) datasets. These datasets present complex scenarios with various defect types characterized by subtle differences from normal samples. DifferNet demonstrated high detection accuracy, achieving superior area under the Receiver Operating Characteristic (ROC) curve values compared to traditional anomaly detection techniques such as One-Class SVM, 1-Nearest Neighbor, and other deep anomaly detection models.

Results highlight that DifferNet not only provides robust anomaly detection performance with limited data – as few as 16 non-defective samples – but also surpasses the state-of-the-art with larger datasets. The method also offers the capability to localize defects, pinpointing anomalous segments of the input with reasonable precision.

Implications and Future Directions

The implications of DifferNet are significant for industrial quality control processes. Its requirement for minimal training data and ability to adapt to various defect types without additional supervision present a valuable tool for continuous quality assurance in manufacturing lines. This capacity aligns well with the unpredictable nature of production defects, making DifferNet a practical solution across several industries.

Theoretically, this work illustrates the potency of normalizing flows in handling complex, high-dimensional data distributions when paired with effective feature extraction strategies. The bijective nature of normalizing flows offers promising avenues for anomaly detection applications, potentially extending beyond visual data to time-series and multi-modal data integration.

Future work can explore optimizing the method for video data anomaly detection, enhancing real-time surveillance and monitoring systems in manufacturing and other domains. Additionally, further exploration of different feature extractor architectures and transformations could refine the robustness and generalization of the approach, allowing for broader applicability and improved localization accuracy.

Overall, the paper presents a compelling advancement in defect detection using semi-supervised learning principles that can significantly impact both theoretical research and practical applications in the field of computer vision and AI.