Deep-learning-based out-of-distribution data detection in visual inspection images

Paper i proceeding, 2023

Within quality critical industries, e.g. aerospace, quality control with non-destructive evaluation (NDE) is essential. The surface quality is often important and e.g. visual inspection is often applied. Part of the inspection is the data interpretation, not easily made automatic for critical products. Recent studies on the automatization have indicated promising results utilizing deep-learning-based artificial intelligence. However, many such algorithms are known to be overconfident when subjected to unexpected input (e.g. new/rare material defects) far from the training dataset, so-called out-of-distribution (OOD) data. We claim that safe computer-based interpretation of NDE data within quality critical applications, must respond sensible also to OOD data. A sensible response could be that the algorithms identify such OOD data and forward it to a human for further analysis. Such an OOD detector could facilitate a human-machine collaboration in a NDE 4.0 vision. In this work we have explored if a recently proposed (for industrial X-ray images) auto-encoder-based approach can be utilized as OOD detector (one-class classifier) for visual inspection data. The model is trained in an unsupervised manner on accepted input to reconstruct it at high precision. Simultaneously it is trained to remove synthetically added defect indications to generate a clean image patch, similar to denoising-auto-enoders. The difference between the input and reconstructed input is analyzed for OOD detection. We train and test the algorithm on a publicly available visual inspection dataset with surface defects. We achieve true positive rates at 0.90 with true negative rates at 0.99 and demonstrate detection of OOD data.