(FL)the defect will substantially deteriorate the quality of the rolled steel strips which lead to poor quality products. It’s essential to adopt a timely and precise strip steel surface defect detection technique. To make this technique available, we combine a novel technique that uses an enhanced ResNet-101 and Federated Learning (FL). By applying ResNet-101 for feature extraction of the images of rolled steel strips, it helps improve the original version based on the FCN network in accuracy. (FCL)residual units of ResNet-101 do not share the same parameters, so the output dimensions of each unit will be inconsistent. Thus the number of nodes required for a fully connected layer is different across each unit. Hence, SPP layers are used to transform outputs of different residual units into identical dimensions in order to form an FC layer later on. Moreover, QPSO algorithm is applied to solve the problem of tuning the parameters that can efficiently converge to the solution of BP learning. The ResNet-101 network model size. In practice, the number of defect samples is usually insufficient and imbalanced for each detection tool. To address the above issue in traditional federated learning methods, we introduce a local data-sharing-based optimization scheme. The method’s performance is measured experimentally using feature visualization techniques and a normalized confusion matrix. The proposed approach can achieve a detection accuracy of up to 98. 83%, and a detection time as low as 0. 04 s.