Deep convolutional neural network-based automated optical inspection for aerospace components

• Automatic Optical Inspection for Aerospace Composite Components. • Create Aerospace Composite Material Image Dataset (ACMID). • Augmented ACMID using Deep Convolutional Generative Adversarial Networks. • Proposed hybrid model combining features of fine-tuned DCNN and classical ML. • Detection accuracy close to that of publicly available datasets. Aerospace manufacturing industry uses composite materials extensively as structural components in civilian and military aircraft. To ensure the quality of the product and high reliability, manual inspection and traditional automatic optical inspection have been employed to identify the defects throughout production and maintenance. These inspection techniques have several limitations such as tedious, time-consuming, inconsistent, subjective, labor intensive, and not cost effective. To overcome these limitations, this research proposes a domain-specific, novel, and efficient Automated Optical Inspection (AOI) method. To our knowledge, this is the first work to combine Generative Adversarial Network (GAN)-augmented data generation with a hybrid Deep Convolutional Neural Network (DCNN) and classical Machine Learning (ML) model to inspect the defects of aerospace components automatically using composite images collected in the Aerospace Composite Material Image Dataset (ACMID). First, two classical ML models, Support Vector Machine and Random Forest are trained and utilized on the dataset, followed by two enhanced DCNN-based architectures. Third, the paper proposes an efficient deep learning AOI method that combines the features of DCNN and classical ML models which is tested on the ACMID dataset. Given the limited samples in the dataset, the research also includes options to address scarce and imbalanced datasets and data augmentation using Deep Convolutional Generative Adversarial Networks (DCGAN). To assess the quality of the aerospace composite components, all the models are trained and tested on diverse set of experimental setups. Experimental results show that the best detection accuracy is given by the proposed method and can reach up to 99.68 percent.