Using elastomer deformation to measure object surface features in tactile sensing is effective, as it captures microscale deformations through densely arranged optical imaging sensors that detect subtle data variations. To enable continuous contact recognition, elastomers are crafted with curved surfaces to adjust to changes in the contact area. However, this design leads to uneven deformations, distorting tactile images and inaccurately reflecting the true elastomer deformations. This inconsistency considerably reduces the utility of the tactile data. In this work, we propose a cyclic fusion strategy for vision-based tactile sensing for precise contact data extraction and shape feature integration at the pixel level. Utilizing frequency domain fusion, the system merges topography as indicated by elastomer deformation, enhancing information content by over 40% and preserving structural consistency. Further, this system could effectively extract and summarize micro-scale contact features, using neural networks to achieve a detection mAP of 90.90% and classification accuracy of 99.83%. Using this strategy, the measurement minimizes data interference, accurately depicting object morphology on tactile images and enhancing tactile sensation restoration.
翻译:暂无翻译