In this thesis we propose a new deflectometry-based anomaly detection approach applicable to reflective surfaces. Classic deflectometry methods detect surface anomalies by performing partial 3D surface reconstruction and differencing it with a pre-recorded reference model of the observed object. Most of these methods require projection of several patterns and require accurate calibration between the pattern projector, camera and the inspected object. In contrast, our anomaly detection approach is defined as a semantic segmentation problem and performs pixel-wise anomaly classification. We utilize the power of deep models for this purpose. Since the proposed method can be trained on annotated anomaly examples, reference objects are not needed, the detection is fast, requires only a single pattern projection and does not require accurate calibration. Furthermore, a robust method for anomaly localization from the segmentation mask is proposed, capable of extracting partially overlapping detections. Preliminary analysis and experimental evaluation were performed to justify the architecture and hyper-parameters of our deep semantic segmentation model. The final model was trained and evaluated on the problem of dent detection in car roofs, where a significant improvement over the base method has been shown. Our model achieves a precision of 0.88, recall of 0.88 and F-score of 0.88 on test data, which represents a nearly 50% improvement over the base method.
|
