Loading...

Image Anomaly Detection based on Deep Learning

Lagzian, Arash | 2023

32 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56380 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Beigy, Hamid
  7. Abstract:
  8. The detection of unusual events, which is called abnormality, is very important in various fields such as: industry, medicine, art, and agriculture, and has applications such as food quality detection, inconsistency detection in work environments, disease detection in medical images, and artefact detection. It has the art of counterfeiting and detection of unhealthy agricultural products, and by detecting the abnormality, the damages caused by the abnormality can be reduced. There are many challenges when detecting anomalies in images, which can be mentioned: the input image is not rich enough to learn a suitable representation, there are not enough samples to learn the model, the high computational cost of the existing methods and inappropriate accuracies. In recent studies by researchers, among the available methods; the use of deep learning and deep neural networks for anomaly detection have been welcomed in various supervised, unsupervised and reinforcement learning approaches. In this thesis, by examining different approaches to solving the problem of anomaly detection, it has been categorized and expressed the advantages and disadvantages of each of the previous researches in this field. Then, by focusing on the approach of compression and knowledge distillation and specifically the architecture of the teacher-student, this problem has been investigated from different aspects and by expressing four main ideas, it has been tried to improve the shortcomings of the previous methods. In the first and second ideas, by paying special attention to the student model and injecting noise into the input and middle layers of the student model, an attempt has been made to increase the efficiency of this model, the results of which report up to 3 percentage improvement. In the third and fourth ideas, with Focusing on the process of knowledge distillation and providing methods based on summation and attention mechanism, an attempt has been made to improve the performance of previous methods of knowledge distillation, the results of which report up to 2 percentage improvement on the standard benchmark dataset
  9. Keywords:
  10. Anomaly Detection ; Deep Learning ; Deep Neural Networks ; Knowledge Distillation ; Reinforcement Learning

 Digital Object List

 Bookmark

...see more