Data augmentation framework for improved classification in object detectors

Deep Learning techniques have been classified as a significant advancement for data-driven industries such as electrical machine manufacturing. It has been used successfully for quality inspection in various manufacturing cases, enabling automated product inspection. However, a massive bottleneck in deploying deep learning for quality inspection in manufacturing operations is the unavailability of training data required to develop an effective model. Data augmentation techniques offer a solution by increasing the size and diversity of the training dataset. In recent years, studies have shown an improvement in the performance of object detectors through image augmentation. However, there is a gap in knowledge regarding the choice of an image augmentation technique and the reasoning behind the choice. This study proposes a framework for choosing an appropriate augmentation method in various scenarios. It was achieved by investigating pixel-level transformations in the context of data augmentation to enhance the training performance of Quality Inspection (QI) models in scenarios with extremely limited data. Firstly, the effects of four pixel-level transformations on YOLOv7 and SSD real-time detectors are analysed. Secondly, the augmentation effect transferability across dissimilar datasets is discussed. Lastly, a framework is proposed for choosing the appropriate augmentation method in various scenarios. The results show that the proposed framework provides consistent results with an average improvement of up to 35% in meanAP and a reduction of 50% in the classification loss for the electrical wires dataset.