Work

FDDA: A Feature Disentangling Domain Adaptation Method for Robust Industrial Defect Detection

Year: 2026

Type: article

Abstract: Robust visual defect detection is crucial for intelligent manufacturing. However, changes in production conditions or image acquisition systems may shift feature distributions, leading to performance degradation. Since the majority of samples are normal, new scenes cannot provide adequate defects for training a new model in the short term. To ensure the robustness of the detection system, a Feature Disentangling Domain Adaptation method (FDDA) is proposed. First, we propose Feature Disentanglement Pre-training (FDP), which disentangles image features into domain-specific and domain-invariant components via domain discrimination and alignment losses. Then, we propose a Domain Prompt Adaptation (DPA), where the main parameters are frozen to preserve the disentanglement and feature extraction capabilities, and the domain-specific modules and the detection head are updated to align the distribution. To explicitly model global and instance-level domain-specific features, we propose a Transformer Encoder Prompt Generation (TEPG) module and a Transformer Decoder Channel Separation (TDCS) module. We construct a detector based on the aforementioned methods and conduct extensive experiments on three multi-domain defect detection datasets. Experiments show that when the number of images in the target domain is limited, FDDA achieves an average improvement of 4.4 AP points over the full fine-tuning and 7.5 AP points improvement compared to the state-of-the-art domain adaptation method. FDDA enables the defect detection system to achieve efficient and robust adaptation under domain shift. Our code is publicly available athttps://github.com/wangzhaoyang-508/FDDANote to Practitioners—At present, the use of AI models for automated industrial defect detection is a key direction in engineering. However, many models suffer from poor robustness, and their detection performance is severely degraded due to changes in production lines or acquisition equipment. The underlying reason is that the image features extracted by these models experience domain shift. To address this problem, we propose FDDA, a two-step Feature Disentangling Domain Adaptation method. This method first pre-trains the model to disentangle image features. When domain shift occurs, the parameters are frozen to retain the ability to disentangle and extract the domain-invariant features, while efficiently updating domain-specific features. Using FDDA can effectively mitigate the performance degradation caused by domain shift and delivers state-of-the-art results. Additionally, we have developed a multi-line automated inspection system based on FDDA and implemented it at the Yingli photovoltaic (PV) cell production factory in China. The application demonstrates that our method enhances the robustness of the defect detection system, reduces costs, and improves automation levels on the production floor.

Source: IEEE Transactions on Automation Science and Engineering

DOI: 10.1109/tase.2026.3726813

Authors: Zhaoyang Wang, Haiyong Chen, Binyi Su, Shijie Wang, Amal Altaezi

Institution: Hebei University of Technology

Language: English

Topic: Industrial Vision Systems and Defect Detection

Subfield: Industrial and Manufacturing Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,645