Leveraging Transfer Learning for Efficient Surface Defect Detection on Metallic Components

Surface defect detection is a key quality control procedure in mechanical engineering, where precise and prompt identification of faults is crucial for sustaining product quality and minimizing production expenses. Conventional manual inspection techniques are labor-intensive and susceptible to human mistake, rendering them inadequate for contemporary production operations. This work examines the efficacy of feature-based transfer learning for detecting surface defects in mechanical engineering components, specifically Ball Screw Drives (BSD) and Metallic Semi-finished Products (SEV). The suggested method utilizes the discriminative features acquired from pre-trained Convolutional Neural Network (CNN) models to enhance defect detection precision and efficacy. A comparison is made between the performance of lightweight and heavyweight CNN architectures when paired with several classifiers, including Support Vector Machines (SVM), Logistic Regression (LR), and K-Nearest Neighbors (KNN). The experimental findings on a dataset comprising BSD and SEV surface images illustrate the efficacy of the suggested method in accurately and effectively detecting faults. The paper elucidates the benefits of feature-based transfer learning compared to conventional approaches and examines its applicability in practical mechanical engineering contexts. The results advance the creation of automated surface flaw detection systems that enhance product quality and decrease manufacturing expenses.