Optical Flow-Based Depth Prediction and Defect Detection Using TacRoller with MOGA-Enhanced U-Net

Vision-based tactile sensors (VBTS) offer high-resolution surface inspection, but current methods often suffer from high computational costs, sensitivity to environmental noise, and poor generalization to unseen materials. This paper presents a hybrid framework that addresses these challenges by integrating RAFT-based optical flow for real-time depth estimation with a Multi-Order Gated Aggregation (MOGA)-enhanced U-Net for defect segmentation. Depth changes between consecutive tactile frames are estimated using geometric constraints from the pinhole camera model, removing the need for explicit 3D reconstruction and reducing runtime by 63% compared to GelSlim. For defect segmentation, MOGA-U-Net combines gated convolutions and attention mechanisms to suppress imaging artifacts such as dust and enhance multi-scale defect features. Evaluated on a composite material dataset with 2,744 defect samples, the proposed method achieves 98.9% detection accuracy at 30 FPS. Integrated with TacRoller—a modular, rolling tactile sensor with a scanning speed of 11.73 mm/s—this system enables efficient, large-area inspection and moves VBTS technology closer to practical deployment in industrial settings.