Source-linked AI summary
Contrastive Knowledge Distillation for Anomaly Detection in Multi-Illumination/Focus Display Images
Jihyun Lee, Hangil Park, Yongmin Seo, Taewon Min, Joodong Yun, Jaewon Kim, Tae-Kyun Kim
TL;DR
The paper targets minute display-surface defects that are difficult to detect from RGB images and normal-only training. It combines contrastive knowledge distillation with multi-channel blending, and reports superior AUROC and accuracy against competitive methods on MMdAD.
Problem
Minute display-surface defects are difficult to detect from RGB images, while most existing anomaly-detection methods use only normal data.
Method
The method extends MKD with Multiresolution Contrastive Distillation and a blending module that aggregates multi-channel information for the framework's three-channel input.
Results
The proposed method outperforms competitive state-of-the-art methods on MMdAD in AUROC and accuracy.
Takeaways & Limitations
Contrastive distillation and multi-channel blending provide a knowledge-distillation approach for anomaly detection in multi-illumination and multi-focus display images.
Abstract
from arXiv · showhide
In this paper, we tackle automatic anomaly detection in multi-illumination and multi-focus display images. The minute defects on the display surface are hard to spot out in RGB images and by a model trained with only normal data. To address this, we propose a novel contrastive learning scheme for knowledge distillation-based anomaly detection. In our framework, Multiresolution Knowledge Distillation (MKD) is adopted as a baseline, which operates by measuring feature similarities between the teacher and student networks. Based on MKD, we propose a novel contrastive learning method, namely Multiresolution Contrastive Distillation (MCD), which does not require positive/negative pairs with an anchor but operates by pulling/pushing the distance between the teacher and student features. Furthermore, we propose the blending module that transforms and aggregate multi-channel information to the three-channel input layer of MCD. Our proposed method significantly outperforms competitive state-of-the-art methods in both AUROC and accuracy metrics on the collected Multi-illumination and Multi-focus display image dataset for Anomaly Detection (MMdAD).
1 Introduction
The paper addresses difficult display-surface anomaly detection by using multi-illumination and multi-focus images, a contrastive distillation scheme, and multi-channel blending. It extends MKD to exploit limited abnormal data and reports state-of-the-art results on MMdAD.
- Motivation: Display-surface defects are difficult to detect from RGB images because defects are minute and specular reflection complicates image patterns.
- Dataset: The collected MMdAD samples contain 12 channels captured under different illumination and focal-length settings.
- Method: The method extends MKD with Multiresolution Contrastive Distillation, pulling teacher–student features together for normal inputs and apart for abnormal inputs.
- Method: A blending module transforms and aggregates multi-channel information for the three-channel input expected by MKD.
- Results: The proposed method achieves state-of-the-art anomaly detection results on the MMdAD dataset.
2 Related Work on Anomaly Detection
Prior anomaly-detection methods achieve strong benchmark accuracy, but most use only normal data and knowledge-distillation methods have been particularly effective on industrial defects.
- Most existing anomaly-detection methods are trained unsupervised using only normal data.
- Knowledge-distillation methods have achieved the highest accuracies on industrial defect benchmarks.
- Anomaly detection is closely related to out-of-distribution detection, open-set recognition, and one-class learning.
3 Multi-Illumination and Multi-Focus Display Image Dataset for Anomaly Detection (MMdAD)
MMdAD is a 12-channel display-image dataset containing normal and anomalous samples captured across illumination and focal-length conditions. Its normal samples can also contain anomaly-like features, making discrimination challenging.
- Dataset composition: MMdAD contains 134 anomalous and 327 normal display images, each with 12 channels and 150 x 150 pixel resolution.
- Dataset composition: The 12 grayscale channels vary in illumination and focal length and include cropped local regions likely to contain anomalies.
- Dataset challenge: Normal MMdAD samples can contain anomaly-like features, distinguishing the dataset from other public anomaly-detection benchmarks.
4 Methodology
The method extends MKD with contrastive distillation that separates teacher–student feature relationships for normal and abnormal images, while a blending module adapts 12-channel display inputs to MKD’s RGB interface.
- MKD baseline: MKD uses a teacher pretrained on large-scale image datasets and a student trained on normal images to imitate teacher feature vectors.Its multiresolution loss combines feature-value and feature-direction similarities.
- Multiresolution Contrastive Distillation: MCD extends MKD to use abnormal examples by contrasting teacher–student relationships in Euclidean and angular feature spaces.It is designed for settings where both normal and abnormal display images are available during training.
- Multiresolution Contrastive Distillation: The contrastive loss makes student features resemble teacher features for normal inputs but disresemble them for abnormal inputs.Class-dependent gradient inversion implements this distinction, while a maximum operator and α bound the abnormal-loss magnitude to improve training stability.
- Multiresolution Contrastive Distillation: Unlike standard contrastive learning, MCD contrasts teacher–student distances without requiring an anchor-based positive/negative triplet.It uses feature similarity in angular space alongside Euclidean distance to distinguish normal and abnormal samples.
- Blending Module: The blending module converts 12-channel multi-illumination and multi-focus images into latent features, then reconstructs an RGB image for MKD processing.The reconstructed image selectively retains essential information from every input channel for anomaly detection.
5 Experimental Results
Experiments on MMdAD evaluate the method against state-of-the-art baselines, ablate its blending and contrastive components, and visualize its anomaly localization behavior.
- Dataset and metrics: The experiments use two-fold cross-validation on MMdAD and report AUROC and accuracy as evaluation metrics.Each split contains 124 normal and 67 abnormal display images.
- Comparisons with state-of-the-art methods: The method outperforms One-Class SVM, RD4AD, DRA, and MKD in AUROC on MMdAD.Baselines use manually selected channels 0, 2, and 6 because they require three-channel inputs.
- Ablation study: The blending module improves performance by over 4% over MKD in both data splits.The ablation settings without blending use channels 0, 2, and 6.
- Ablation study: Contrastive learning adds a further 2-4% performance improvement over the corresponding settings in both data splits.Figure 3 provides ROC-curve support for the ablation results on dataset split 1.
- Gradient heatmap visualization: The gradient heatmaps show that the model attends to abnormal regions while exploiting multi-channel information.In the heatmaps, red denotes high gradients and blue denotes low gradients.
6 Conclusion
The paper proposes contrastive knowledge distillation for anomaly detection on MMdAD and adds a blending module for multi-channel inputs. The proposed method outperforms competitive state-of-the-art methods on anomaly detection.
- Conclusion: The paper proposes a novel contrastive learning scheme for knowledge distillation-based anomaly detection on MMdAD.
- Conclusion: The blending module transforms and aggregates multi-channel information into the framework's input layer.
- Conclusion: The proposed method outperforms competitive state-of-the-art methods on anomaly detection.