A Parameter Decoupling and Spatial Vision Mamba-enhanced Model for Mini LED Anomaly Detection

Mini Light-Emitting Diode (Mini LED) plays a crucial role in modern display technology, but various defects inevitably arise during its manufacturing process, making accurate detection essential. However, conventional supervised learning models struggle in this scenario due to the continuous emergence of new types of defects and the scarcity of samples for certain defect categories. To address this issue, this paper proposes a novel unsupervised anomaly detection method based on EfficientAD, which theoretically enables the detection of many kinds of defects by learning only the distribution of normal samples. We introduce two key improvements: Firstly, a separation student network that decouples the dual-channel outputs at an earlier stage, allowing independent parameter learning for teacher network features and autoencoder reconstructions, address the domain gap between these two feature spaces. Secondly, a spatial vision Mamba (SVM) module is integrated into the autoencoder, which employs bidirectional Mamba scanning with a gated fusion mechanism to provide high-quality feature enhancement through long-range dependency modeling while maintaining linear complexity. Our method yielded AUROCs of 96.61% (image-level) and 95.75% (pixel-level) on the self-constructed Mini LED dataset, accurately detecting both foreign materials missing chips and colored diode.