Related Experiment Video
Updated: Jun 29, 2025
08:12
A Psychophysics Paradigm for the Collection and Analysis of Similarity Judgments
Published on: March 1, 2022
2.5K
Cosine similarity knowledge distillation for surface anomaly detection
Siyu Sheng1, Junfeng Jing2,3, Zhen Wang4
1College of Electrical and Information, Xi'an Polytechnic University, Xi'an, 710048, China.
Scientific Reports
|April 8, 2024
Summary
We introduce Cosine Similarity Knowledge Distillation (CSKD) for precise surface anomaly detection. This method enhances learning and outperforms existing unsupervised anomaly detection algorithms on complex backgrounds.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Current knowledge distillation (KD) methods for anomaly detection struggle with complex textures due to representation discrepancies.
- Existing approaches often use smaller student networks or reverse distillation, limiting precision.
Purpose of the Study:
- To propose a novel Cosine Similarity Knowledge Distillation (CSKD) paradigm for enhanced surface anomaly detection and localization.
- To improve the performance of teacher-student models in identifying anomalies within complex backgrounds.
Main Methods:
- Implemented Cosine Similarity Knowledge Distillation (CSKD) using identical deep teacher and student encoders.
- Introduced Attention One-Class Embedding (AOCE) within the student network to boost learning and mitigate teacher-student response similarity issues in anomalous regions.
- Developed an adaptive optimal model selection method based on class-specific hard-coded epochs.
Main Results:
- Achieved 99.2% image-level AUROC and 98.2%/94.7% pixel-level AUROC/PRO on the MVTec dataset.
- Demonstrated superior performance compared to existing unsupervised anomaly detection algorithms.
- Validated effectiveness on the DAGM dataset and other one-class anomaly detection benchmarks.
Conclusions:
- CSKD offers a significant advancement in unsupervised surface anomaly detection and localization.
- The proposed AOCE and adaptive model selection contribute to robust and precise anomaly identification.
- The method shows strong generalization capabilities across diverse datasets and anomaly detection tasks.
Related Concept Videos
01:29Direction Cosines of a Vector
503
Direction cosines, which help describe the orientation of a vector with respect to the coordinate axes, are an essential concept in the field of vector calculus. Consider vector A that is expressed in terms of the Cartesian vector form using i, j, and k unit vectors. The magnitude of vector A is defined as the square root of the sum of the squares of its components. The direction of this vector with respect to the x, y, and z axes is defined by the coordinate direction angles α, β, and γ,...
503
01:18Aliasing
133
Accurate signal sampling and reconstruction are crucial in various signal-processing applications. A time-domain signal's spectrum can be revealed using its Fourier transform. When this signal is sampled at a specific frequency, it results in multiple scaled replicas of the original spectrum in the frequency domain. The spacing of these replicas is determined by the sampling frequency.
If the sampling frequency is below the Nyquist rate, these replicas overlap, preventing the original...
If the sampling frequency is below the Nyquist rate, these replicas overlap, preventing the original...
133
01:24Area Computation by the Alternative Coordinate Method
52
The alternative coordinate method, also known as the Shoelace Formula, is a technique for determining the area of a traverse using Cartesian coordinates. This method relies on the sequential arrangement of x and y coordinates for each point of the shape, ensuring accuracy and ease of application.In this approach, each corner's x and y coordinates are listed as fractions, with the x-coordinate as the numerator and the y-coordinate as the denominator. These coordinates are arranged sequentially...
52