A Multimodal Unsupervised Coating Defect Detection Method Based on Dual-Branch Hybrid CNN-Mamba U-Net

Ensuring the integrity and functionality of coatings is vital in modern industrial processes. However, coating defect detection faces significant challenges, mainly due to the lack of standardized datasets and the unpredictable diversity of defects. Moreover, most studies focus on unimodal detection methods. To address these challenges, we develop a standard multimodal coating defect dataset (MCD-AD) and propose a multimodal unsupervised coating defect detection method based on dual-branch hybrid convolutional neural network (CNN)-Mamba U-Net (DCMUNet). First, to tackle the lack of a coating depth anomaly dataset, we propose a depth anomaly generation strategy. Then, in the reconstruction network, we combine a CNN-based local feature encoder with decoder based on hybrid CNN-Mamba (HCM) basis block for joint feature extraction to capture local-global context information. Additionally, we introduce a multiscale feature enhancement (MSFE) module and an adaptive feature weighting (AFW) module in the 2-D discrimination network to integrate and refine features. Finally, a multimodal feature fusion module is designed to facilitate the fusion of RGB and depth features. Furthermore, depthwise separable convolution is employed to achieve a lightweight design for DCMUNet. We conduct extensive experiments on MCD-AD, MVTec 2D-AD, and MVTec 3D-AD datasets. The results demonstrate that our method outperforms the current state-of-the-art unimodal and multimodal approaches, achieving refined segmentation of coating defects while also exhibiting robustness across various industrial application scenarios. The code and MCD-AD dataset are available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/TK941025/DCMUNet</uri>.