View Proposal

---

Proposer

Wei Pang

Title

Open Set Recognition for Industrial Anomaly Detection

Goal

Description

This research project aims to explore and apply Open Set Recognition (OSR) techniques to the field of Industrial Anomaly Detection (IAD). The goal is to address the limitations of existing industrial quality inspection systems when confronted with unknown and novel defects. Traditional anomaly detection methods typically operate under a "closed-set" assumption, meaning all possible defect types are known during the training phase. However, in real-world industrial production environments, due to the complexity and dynamic nature of manufacturing processes, new and unforeseen defect types constantly emerge. This project will tackle this challenge by researching and developing a deep learning model capable of effectively identifying known defects while simultaneously rejecting unknown defect types.

Resources

Background

Current industrial quality inspection systems largely rely on supervised or semi-supervised learning methods. The core principle of these approaches is the "closed-set" assumption, meaning all potential defect types must be known during the training phase. However, this assumption often proves untenable in real, dynamic production environments, leading to several significant limitations: • High-Cost and Lagging Manual Annotation: Building and updating datasets heavily depends on manual annotation. This process is not only costly and time-consuming but also suffers from significant latency. When a new defect type appears on the production line, it takes time to collect sufficient samples, perform manual identification and precise annotation, and finally incorporate them into the model's retraining process. During this unavoidable "window of vulnerability," the automated system is entirely "blind" to such new defects. • Continuous Emergence of Novel Defects: Industrial production is a dynamic process. Batch variations in raw materials, natural wear and tear of equipment, subtle adjustments in process parameters, and even environmental changes can all give rise to entirely new, unforeseen defect types at any moment. Waiting for these defects to be discovered and annotated means that a batch of potentially defective products has already been produced or even released. • Limitations of Traditional Models: Traditional anomaly detection systems, when faced with defects falling within this "window of vulnerability" or entirely unknown defects, exhibit two typical failure modes: either they mistakenly classify them as "normal" products, leading to missed detections; or they forcibly categorize them into a known, feature-similar defect class, leading to misclassifications. Both scenarios directly result in quality control failures, posing severe economic losses and reputational risks to enterprises. Therefore, to overcome these bottlenecks, it is crucial to research a detection paradigm that does not entirely rely on historical labelled data. Open Set Recognition (OSR) technology offers a solution by empowering systems with the ability to "recognize what they don't know." This means accurately identifying known defects while effectively detecting and rejecting any unknown defect samples that do not belong to the known classes. Applying OSR to industrial anomaly detection directly addresses the challenges posed by the limitations of manual annotation and the dynamic nature of production. It enables the system to shift from "passively awaiting annotation" to "actively discovering the unknown," holding immense theoretical value and urgent practical demand for enhancing the intelligence, robustness, and real-world adaptability of automated quality inspection systems.

Url

Difficulty Level

High

Ethical Approval

None

Number Of Students

1

Supervisor

Wei Pang

Keywords

deep learning, open set recognition, anomaly detection

Degrees

Bachelor of Science in Computer Science
Master of Science in Artificial Intelligence
Master of Science in Computing (2 Years)
Master of Science in Data Science
Bachelor of Science in Computing Science
Bachelor of Engineering in Robotics
BSc Data Sciences