Automated Visual Inspection for Defects Using Data Augmentation for Deep Learning-Based Image Classification

This article explores automated visual inspection for classifying defects in industrial products. Visual inspection to detect defective samples poses a significant challenge in industries due to the inherent variations present. Gathering an adequate amount of data for defect detection and classification using deep learning is particularly challenging in industrial automation due to the scarcity of defective samples. To address this issue, various data augmentation methods have been proposed, including geometric transformations, image fusion, variational autoencoders (VAE), and Generative Adversarial Networks (GANs). Convolutional Neural Network (CNN) model for defect classification was developed and evaluated using a dataset of electrical pin connectors augmented by image processing techniques and variants of conventional GANs, namely Deep Convolutional Generative Adversarial Network (DCGAN) and Wasserstein Generative Adversarial Network (WGAN). The developed model achieved an accuracy of over 90% on all three uniquely augmented datasets, representing an improvement in performance compared to the model's performance on the original dataset. However, it is worth noting that GAN-based methods were more time-consuming and computationally expensive. Hence, when considering data augmentation methods for implementation, it is advisable to first explore image processing techniques. If these do not yield satisfactory results, a combination of image processing and GAN-based methods may be considered.