Federated learning-based defect localization on metal surfaces using ensembled pretrained deep neural networks with feature fusion

Abstract This work proposes an effective approach towards effective classification of metallic surface defects by using feature extraction, feature reduction, deep neural networks and federated learning. The process starts with using the GC10-DET dataset which consists of 3,570 grayscale images of ten different types of defects on steel sheet namely punching, weld line, and oil spot among others. Firstly, the images are preprocessed and normalized where the dimensions of the images are resized to a certain size and also the images may be augmented to have similar format. The feature extraction is done by five Efficients models—B0 to B4, and this results in extracting 7,296 features for each image. Linear Discriminant Analysis (LDA) is then used, which projects feature space into nine significant features to enhance its readiness for classification. Classification is being performed through Deep Neural Network (DNN), Long Short Term Memory (LSTM), and Bi-directional LSTM (Bi-LSTM). Among these, EfficientNetB2 combined with LDA demonstrates superior performance, achieving a validation accuracy of 88.7446%, precision of 90.4545%, and recall of 86.1472%. Federated learning is employed to preserve data privacy while enabling collaborative model training across four clients, representing different data holders. The system shows impressive performance in both Independent and Identically Distributed (IID) and Non-IID data scenarios, with clients achieving up to 100% accuracy, precision, and recall under IID conditions, and rapid convergence to near-perfect accuracy under Non-IID conditions. These results highlight the methodology’s effectiveness in real-world industrial applications, where high accuracy and data privacy are paramount.