SSCL: Semi-supervised Contrastive Learning for Industrial Anomaly Detection

Anomaly detection is an important machine learning task that aims to identify data points that are inconsistent with normal data patterns. In real-world scenarios, it is common to have access to some labeled and unlabeled samples that are known to be either normal or anomalous. To make full use of both types of data, we propose a semi-supervised contrastive learning method that combines self-supervised contrastive learning and supervised contrastive learning, forming a new framework: SSCL. Our method can learn a data representation that can distinguish between normal and anomalous data patterns, based on limited labeled data and abundant unlabeled data. We evaluate our method on multiple benchmark datasets, including MNIST, CIFAR-10 and industrial anomaly detection MVtec, STC. The experimental results show that our method achieves superior performance on all datasets compared to existing state-of-the-art methods.