LPFSTNet: A lightweight and parameter-free head attention-based student–teacher network for fast 3D industrial anomaly detection
Yahui Cheng, Junchao Chen, Guojun Wen, Xianhua Tan, Xingyue Liu · Neurocomputing · 2025
The emerging three-dimensional (3D) anomaly detection networks have achieved remarkable detection accuracy in industrial applications. However, they always occupy excessive computing resources and exhibit bad real-time performance. Herein, a lightweight and parameter-free head attention-based student–teacher network (LPFSTNet) is demonstrated for unsupervised 3D anomaly detection. Firstly, a lightweight student network mainly containing simplified convolution operations is constructed, helping to reduce the parameter numbers while increasing the distinction between the student and teacher networks, thus enhancing generalization capability of the model. Then, a novel parameter-free head attention mechanism is proposed to regulate the output of the model, contributing to a higher detection accuracy. This head attention module can be directly integrated into the model without any training and adds no additional inference time. The experimental results validate that LPFSTNet achieves 94.7% image-level area under the receiver operating characteristic curve (I-AUROC) and 38.5 frames per second (FPS) on MVTec3D dataset, surpassing other mainstream 3D anomaly detection counterparts. Importantly, the model size is decreased by 74% compared to the baseline. Moreover, the experiment implemented on our own dataset verifies that our LPFSTNet model is robust to other dataset. The code and dataset are available at https://github.com/YahuiCheng/LPFSTNet .