Self-supervised pairing image clustering for automated quality control

In manufacturing, artificial intelligence is attracting widespread attention to maximize industrial productivity. Image clustering, as a fundamental research direction in unsupervised learning, has been applied in various fields. Since no label information is demanded in clustering, a preliminary analysis of the unlabeled data can be done while saving lots of manpower. In this paper, we propose a novel Self-supervised Pairing Image Clustering network. It predicts clustering results in an end-to-end pair classification network, which is trained excluding any label information. Specifically, we devised a self-supervised pairing module that is able to accurately and efficiently create both types of pairs as training data after exploiting the pair distribution. Cooperating with the pair classification loss function, we added two regularization terms to ensure the clustering results to be unambiguous and close to the real data distribution. In the experiments, our method exceeds most of the state-of-art benchmarks in the publicly available manufacturing datasets—NEU and DAGM. Besides, the method also demonstrates an excellent generalization capability on general public datasets including MNIST, Omniglot, CIFAR-10, and CIFAR-100.