Transfer Learning Using Cluster Centers for Surface Defect Identification of Piston Rods in Pneumatic Cylinders

Surface defect identification is essential for piston rods to ensure leakproofness, non-pollution, and control precision of pneumatic cylinders. However, due to environmental conditions, the distribution discrepancy between surface images is unavoidable. Transfer learning (TL) methods can be adopted to reduce this discrepancy, but they utilized the class center to represent samples from one class for measuring the distribution discrepancy, which can not reflect the dispersion of samples. To address this problem, this article considers samples within one class dispersing into multiple clusters, and then proposes TL using cluster centers (TLCC). First, to determine cluster centers within each class, TLCC applies the $K$ -means algorithm to divide samples within each class into $K$ clusters. Second, noticing that using all cluster centers to represent samples from one class can effectively reflect the dispersion of samples, TLCC establishes cluster centers discrepancy (CCD) for measuring the distribution discrepancy between the source and target domains. CDD is designed on the distance between each pair of cluster centers within each class. Next, to reduce the distribution discrepancy, the optimization objective of TLCC is established on minimizing CCD. Finally, sufficient results demonstrate that TLCC outperforms the state-of-the-art TL methods, where the average accuracies of TLCC on the benchmark dataset [CMU pose, illumination, and expression (CMU-PIE)] and our surface image of piston rods used in pneumatic cylinders (SIPRPC) dataset are 85.3% and 98.8%.