Patch-Masked Visual Inspection via Parallel Deep Tensor Factorization in Industrial Internet of Things

Visual inspection is an effective approach for anomaly detection in the industrial Internet of Things. The inpainting-based strategy is widely adopted for industrial visual inspection. However, the training process of the semantic-based inpainting model requires heavy resource consumption due to its complex structure. Additionally, the inspection performance of the model degrades when it encounters an out-of-scope image. Image can be naturally deemed as a tensor with the global low-rank property. Therefore, we propose a data structure-based strategy to implement effective inpainting in this paper. Inspired by the tensor-tensor product, a patch-masked deep tensor factorization model is constructed to reconstruct the intentionally masked region. This model has a simplified structure with three feed-forward neural subnets sharing one trainable low-dimensional input. Besides, Laplacian regularization is imposed to improve the recovery accuracy based on the local smoothness in images. Above all, we apply a parallel approach to implement patch-masked visual inspection by comparing the structural similarity between the recovered and original patches. It is independent of the dataset and works as a sample-related pattern. Experiments conducted on the MVTec dataset demonstrate that, although our model has lower inspection resolution than semantic-based inpainting models, it has much higher training efficiency, which makes it more beneficial for practical deployment.