Exploring the Newly Introduced Data Augmentation Technique in TensorFlow Library

Libraries such as TensorFlow provide various tools for implementing live data augmentation methods. However, in practical applications, common situations such as objects arriving at different angles can negatively impact system performance. For such factors, there is no dedicated live data augmentation method in the TensorFlow library. In this study, a new live data augmentation method that can perform real-time random rotation operations in the user-specified angle and axis range has been developed to overcome this deficiency. This method, which is integrated into the TensorFlow library, is much more flexible than the existing method, which can only perform a 90° rotation on the X-Y axis. Artificial intelligence training was performed on standard (existing method) and modified (new method) model master data by labeling 2800 objects on 1060 images of parts obtained from a private automotive manufacturing facility. The training results were tested for five different scenarios on an additional 115 image data collected from the production facility at different times. It was found that the model trained with the new method outperformed the model trained with the existing method by up to 2.6%. This study will significantly contribute to the detection of moving parts with higher accuracy, especially in vision-based systems.