Ensemble Learning Based Defect Detection Systems for Industrial Applications

Abstract

The efficiency and effectiveness of conventional human inspection techniques for surface defects are very low in most industrial products. The modern industry requires immediate identification of the non-conforming parts. Human inspectors cannot be relied upon for this all the time. Although these processing techniques will help overcome some of the limitations, they don't do well with the complex textures, noise, and light variations. The paper aims to fill in this gap by proposing deep learning techniques for better defect detection and classification. The developed model, on training with various surface textures datasets, presented excellent performance in detecting the cracks, patches, inclusions, rolled edges, pitting, and scratching defects. By adding ResNet50 and MobileNetV2 additive into an ensemble framework model, the detection of the surface problem became more accurate and reliable. This advanced method can potentially be helpful to the industry both in diagnosis and classification.