Noise Fusion-based Distillation Learning
for Anomaly Detection in Complex Industrial Environments

Jiawen Yu1, Jieji Ren2, Yang Chang1, Qiaojun Yu2,3, Xuan Tong1, Boyang Wang1,
Yan Song1, You Li1, Xinji Mai1, Wenqiang Zhang1*
1College of Intelligent Robotics and Advanced Manufacturing, Fudan University, Shanghai, China
2Shanghai Jiao Tong University, Shanghai, China
3Shanghai AI Laboratory, Shanghai, China
Video Code

Video

Abstract

Anomaly detection and localization in automated industrial manufacturing can significantly enhance production efficiency and product quality. Existing methods are capable of detecting surface defects in pre-defined or controlled imaging environments. However, accurately detecting workpiece defects in complex and unstructured industrial environments with varying views, poses and illumination remains challenging. We propose a novel anomaly detection and localization method specifically designed to handle inputs with perturbative patterns. Our approach introduces a new framework based on a collaborative distillation heterogeneous teacher network (HetNet), an adaptive local-global feature fusion module, and a local multivariate Gaussian noise generation module. HetNet can learn to model the complex feature distribution of normal patterns using limited information about local disruptive changes. We conducted extensive experiments on mainstream benchmarks. HetNet demonstrates superior performance with approximately 10% improvement across all evaluation metrics on MSC-AD under industrial conditions, while achieving stateof-the-art results on other datasets, validating its resilience to environmental fluctuations and its capability to enhance the reliability of industrial anomaly detection systems across diverse scenarios. Tests in real-world environments further confirm that HetNet can be effectively integrated into production lines to achieve robust and real-time anomaly detection.

Introduction

In modern manufacturing environments, surface defect detection is crucial for ensuring product quality and structural integrity. As flexible manufacturing systems evolve, traditional fixed-position inspection stations have become inadequate, while robot-based inspection systems introduce new challenges despite their flexibility. HetNet is an innovative unsupervised anomaly detection framework designed to overcome key limitations in robot-based inspection by addressing real production environment challenges including uncontrolled illumination conditions, perspective variations, and magnification differences. Through multi-modal feature extraction mechanisms, HetNet more effectively characterizes the probabilistic distribution of normal patterns, integrating CNN and Transformer networks to simultaneously capture heterogeneous features and local-global contextual information. The framework employs an adaptive local-global feature fusion module that optimizes information flow between complementary feature spaces, implements a collaborative distillation framework executing joint optimization objectives including reconstruction and denoising tasks, and introduces a local multivariate Gaussian noise generation module that effectively expands the decision boundary. HetNet provides a robust solution for industrial automated inspection systems, particularly suitable for automotive, aerospace, and precision machinery sectors where high accuracy and reliability are required.

Pipeline

Our pipeline consists of four modules. The image is input into two heterogeneous encoders to obtain local and global representations. The hybrid feature fusion strategy consists of the ALGF module and the MHF module. ALGF fuses heterogeneous features, facilitating effective interaction between global and local characteristics, and the MHF module combines features from different layers. The collaborative student distills knowledge from the teachers by noise fusion-based distillation learning. The noise for denoising is generated by the LMGN generator. Please check our main paper for more technical details.

Experiments

We validated our method on the MSC-AD dataset, a challenging benchmark for industrial anomaly detection. HetNet outperforms SOTA methods in all categories, improving image-level AUROC by 18.46% over the second-best method and nearly 40% over others. From left to right, we present anomaly localization results for samples under varying illumination, resolution, and positioning conditions. Predictions made by HetNet robustly adapt to input disturbances and exhibit more stable performance across diverse environmental conditions. We conducted ablation experiments to exhibit the effectiveness of each module in HetNet. Additional tests on the MVTec-AD, VisA and MPDD datasets confirmed its generalization ability and robustness.

The extra qualitative comparisons demonstrates HetNet's effectiveness in reducing false positives on normal disruptive patterns.

We deployed HetNet in a real-world inspection setting to detect surface defects on precision metal components.