DefectNet: A Transfer Learning Framework for Multi-Surface Defect Detection for Real-Time Industrial Inspection

Abstract

Surface defects in metallic, ceramic, and electronic components pose significant risks to structural integrity and brand reputation. Conventional quality control relies heavily on manual inspection, a process plagued by subjectivity, high operational costs, and inconsistency. Early automated attempts utilizing handcrafted features such as edge detection and histogram-based texture filters lack the robustness required to navigate the complexities of dynamic industrial environments, including fluctuating illumination and varying defect orientations. To bridge this gap, the study proposes a robust hybrid framework that marries deep transfer learning with optimized machine learning classifiers. The system utilizes the Xception architecture as a deep feature extractor, benefiting from its depthwise separable convolutions to capture intricate surface anomalies with high computational efficiency. These high-dimensional features are subsequently processed through a suite of diverse classifiers, including Stochastic Gradient Descent (SGD), Passive Aggressive Classifier (PAC), Histogram-based Gradient Boosting (HGB), and Quadratic Discriminant Analysis (QDA). Experimental results indicate that the hybrid approach significantly outperforms traditional inspection models in terms of generalization and real-time processing speed. To ensure practical viability, the framework is integrated into a Graphical User Interface (GUI), allowing non-technical personnel to execute seamless image uploads and instant defect classification. This research provides a scalable, high-accuracy solution for predictive quality assurance, minimizing human error and optimizing productivity in modern smart manufacturing.