A Lightweight Reverse Distillation for Texture Anomaly Detection

Reverse Distillation (RD) has achieved effective results in the challenge of unsupervised anomaly detection on the textured surfaces of industrial products. However, original RD has very large numbers of parameters, making it unsuitable for practical deployment. To address this issue, we proposed a lightweight reverse distillation called RD-Lite. Efforts to reduce the model’s weight were made in two parts. First, the depth and the number of feature channels of the model were significantly reduced. Additionally, the student network and the bottleneck layer were reconstructed using depthwise separable convolution modules. The streamlined model maintains excellent anomaly detection performance while keeping the model size at a tiny level. Experiments on benchmark texture datasets show that our RDLite achieved a sample-level AUROC of 98.92%, which is very close to the accuracy of the original RD, while the number of our model’s parameters is only 3.3% of the original one.