Enhancing CNN-based fault detection in seismic data using Hough transform-based geometrical filtering: Application to Opunake dataset

Deep learning has transformed seismic fault detection, with CNNs like FaultNet enabling rapid interpretation. However, these models often misinterpret high-gradient zones, producing large incoherent patches and artifacts instead of precise fault structures due to a lack of geometric constraints. To address this, we propose a Hough-Driven Refinement strategy that applies the Hough transform as a structural filter. This method enhances fault continuity while removing erroneous patches, improving fault detection without retraining the model. Tested on the Opunake 3D seismic dataset, our approach significantly refines FaultNet’s predictions, reducing false positives and improving structural accuracy. Segmentation metrics confirm its effectiveness, demonstrating the potential of integrating deep learning with geometric constraints for more reliable fault interpretation.