Automatic anomaly detection from X-ray images based on autoencoders

Radiography is a common imaging method for non-destructive object inspection. Processing radiographs to detect defects is, however, challenging, especially if the defect type and location are unknown. In other fields, autoencoders (AE) have been largely studied for anomaly detection. While their performance for defect detection has shown promise, many solutions are not demonstrated to be defect-independent. In this work, we introduce a radiography-based anomaly detection method that first computes the AE prediction error from an input radiograph to enhance the detectability of defects. Subsequently, the probability of the imaged object being defective is predicted by a convolutional classification network. Results on simulated data demonstrate that this method succeeds in distinguishing between defective and non-defective objects, with accuracy (which calculates how often the predictions are equal to the labels) >94% and precision (which is the ratio between the true positives and the total predicted positive) >90%, even for defects unseen during the training stage. Our classification network showed high robustness to noise (with signal-to-noise-ratio (SNR) of 20 dB), and accuracy on noisy data >90% and precision >86%. Since our network workflow does not require image registration for defect detection, the proposed solution is therefore independent of the orientation and position of the object during the scan.