TY - GEN
T1 - A Hybrid Deployment Framework for Industrial Quality Control Using GelSight Tactile Sensing
AU - Damodaran, Kaviyashree
AU - Gambale, Antonio
AU - Kerr, Dermot
AU - Coleman, Sonya
PY - 2026/9/30
Y1 - 2026/9/30
N2 - Tactile sensing offers a potential alternative to vision-based defect detection for reflective, glossy, or low-texture materials common in precision manufacturing. This paper addresses the binary classification of surface defects using GelSight tactile imagery, which captures high-resolution geometric deformations through an elastomeric interface. The primary challenge is the significant “domain gap” between natural images used for pre-training deep learning models and the specific tactile patterns of GelSight sensors, which is further exacerbated by the scarcity of large-scale tactile datasets. To address this, we compare a classical Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) pipeline against three ImageNet-pretrained convolutional neural networks (CNNs), including MobileNet V2, ResNet-18 and EfficientNet-B0. The results indicate that in data-constrained tactile perception, hand-engineered features continue to offer unmatched precision and computational efficiency. We recommend a layered deployment strategy for robotic quality control: utilising HOG-SVM for real-time, precision-driven filtering to minimise false positives, and MobileNet V2 for safety-critical scenarios where maximising defect capture is paramount.
AB - Tactile sensing offers a potential alternative to vision-based defect detection for reflective, glossy, or low-texture materials common in precision manufacturing. This paper addresses the binary classification of surface defects using GelSight tactile imagery, which captures high-resolution geometric deformations through an elastomeric interface. The primary challenge is the significant “domain gap” between natural images used for pre-training deep learning models and the specific tactile patterns of GelSight sensors, which is further exacerbated by the scarcity of large-scale tactile datasets. To address this, we compare a classical Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) pipeline against three ImageNet-pretrained convolutional neural networks (CNNs), including MobileNet V2, ResNet-18 and EfficientNet-B0. The results indicate that in data-constrained tactile perception, hand-engineered features continue to offer unmatched precision and computational efficiency. We recommend a layered deployment strategy for robotic quality control: utilising HOG-SVM for real-time, precision-driven filtering to minimise false positives, and MobileNet V2 for safety-critical scenarios where maximising defect capture is paramount.
U2 - 10.48460/MU.MURAL.00021908
DO - 10.48460/MU.MURAL.00021908
M3 - Conference contribution
T3 - Proceedings of the Irish Machine Vision and Image Processing Conference
BT - Proceedings of the 28th Irish Machine Vision and Image Processing Conference (IMVIP 2026)
PB - Irish Pattern Recognition and Classification Society
ER -