Title: Robust learning for fine-grained anomaly detection in a data-rich but label-rare environment

Date: Monday, April 29th

Time: 1:00 PM – 3:00 PM ET

Location: Groseclose 403

Meeting Link: https://gatech.zoom.us/j/96063820519  

Shancong Mou

Industrial Engineering PhD Candidate

H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology

Committee

Dr. Jianjun Shi  (Advisor), School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Jing Li, School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Yajun Mei, School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Kamran Paynabar, School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Hao Yan, School of Computing, Informatics, and Decision Systems Engineering, Arizona State University

Abstract

Image-based anomaly detection is of vital importance across various industrial, environmental, and medical domains. The continuous evolution of sensor technologies has facilitated the collection and storage of high-resolution, multichannel, and multi-view measurement signals, providing detailed product insights. However, annotating defects remains challenging due to the costly labeling process and the infrequent occurrence of defects in modern manufacturing processes, leading to a scarcity of labeled data and a label-rare environment. Moreover, contemporary high-value and safety-critical applications necessitate fine-grained anomaly detection capabilities. Yet, achieving such precision in anomaly detection, in the absence of adequate anomaly supervision signals, presents a big challenge for modern algorithms.

To address these challenges, this doctoral thesis delves into robust learning algorithms designed to achieve robust signal restoration under unknown gross corruptions. In the context of anomaly detection, anomalies are often construed as sparse corruptions within an otherwise clean background signal. This thesis endeavors to extend the capabilities of robust learning algorithms to tackle various real-world challenges.  Specifically, three important topics are investigated in this thesis to address the aforementioned challenges:

(1) Additive Tensor Decomposition: Tensor data is becoming increasingly popular for storing high-resolution, multichannel, and multi-view measurement signals. To accommodate these new data types within such a data-rich environment, traditional robust learning algorithms are being generalized to handle tensor data. As part of this effort, an Additive Tensor Decomposition (ATD) method has been proposed. Within this ATD method, a set of tensor structural priors are defined to effectively model clean background signals, offering a flexible means of integrating engineering knowledge. The proposed method formulates a high-dimensional optimization problem, incorporating regularization terms based on these tensor structural priors, to decompose the tensor into multiple components representing different structural information such as normal background and anomalies.

(2) Compressed Smooth Sparse Decomposition: An immediate challenge stemming from advancements in measurement technology is the handling of storage, transmission, and computation burdens associated with measurement signals. For instance, achieving real-time inspection of minute surface defects often requires the transmission and processing of large volumes of high-resolution images in real time. This considerable amount of high-resolution image data presents significant challenges not only for the speed of decomposition-based image processing algorithms but also for the storage and transmission of the data itself. To address this challenge, a fast and data-efficient method is developed for sparse anomaly detection in high-resolution images with a smooth background. This method, termed Compressed Smooth Sparse Decomposition (CSSD), is a one-step approach that integrates compressive image acquisition with decomposition-based image processing techniques. Theoretical properties regarding the feasibility of applying CSSD are thoroughly examined. Furthermore, the proposed framework is generalized to a tensor signal setting.

(3) Distributional-assumption-free robust learning: In scenarios involving complex signals like computed tomography (CT) scans of human organs or inspection images of manufacturing products with diverse surface patterns, simplistic statistical priors often prove insufficient. Current robust learning methods, which rely on those statistical priors such as low rank and smoothness to reconstruct normal signals and identify abnormal components, frequently fall short in such applications. The abundance of large-scale normal data in high-yield advanced manufacturing applications prompts the incorporation of learned priors into the robust learning framework, thereby modeling signals in a distributional-assumption-free manner. In this thesis, a set of distributional-assumption-free robust learning methods are developed and presented in two consecutive chapters: Part I: Patch Autoencoder Based Deep Image Decomposition (PAEDID) in Chapter 4, and Part II: Robust GAN inversion (RGI) methods in Chapter 5. In these chapters, clean background signals are modeled as deep priors learned from extensive datasets of clean images, encompassing normal product variations. Meanwhile, anomalies are modeled as sparse corruptions to these clean signals. The proposed methods seamlessly integrate robustness from a robust learning framework with the modeling capabilities of deep neural network learned priors, enabling the resolution of challenges that were previously daunting for either deep learning or statistical robust learning in isolation.