Deep CNN-based visual defect detection: Survey of current literature

Abstract

In the past years, the computer vision domain has been profoundly changed by the advent of deep learning algorithms and data science. The defect detection problem is of outmost importance in high-tech industries such as aerospace manufacturing and is extensively employed using automated industrial quality control systems. Defect inspection methods can be mainly grouped into manual inspection, traditional computer vision, and modern computer vision inspection. Initially developed two decades ago, the CNN algorithms recently became popular for solving complex machine vision problems, as big datasets and computationally potent hardware became widely available. Deep learning-based methods form the foundation for modern automatic optical inspection methods and can be grouped based on their network connections into two categories: dense networks and sparse networks. Another method for grouping considers the type of learning: supervised learning used primarily for defect classification and segmentation, and unsupervised learning models, which have the potential to overcome the challenges of supervised models such as labeling images and annotating pixels. In addition, pixel-level based segmentation techniques are considered to cover the state-of-the-art methodologies for the automatic optical inspection. Still, both supervised and unsupervised models pose challenges in regards to model training and attaining the expected detection accuracy. Identified open challenges include algorithmic, application, and data processing challenges. By addressing these challenges, in the future, the demand for automated optical inspection is expected to only grow in both industry practice and academic research.