Object detection, auto-focusing and transfer learning for digital holography of solid composite propellant using efficient neural network

Digital holography has emerged as a powerful tool for various applications. However, applications such as target detection and depth prediction in complex scenarios like composite propellant combustion still face challenges regarding accuracy. To address this issue, we propose a deep neural network based on the vision transformer capable of efficiently and accurately performing multiple tasks in digital holography, including fine-grained target detection and autofocus. Furthermore, we leverage pre-trained models on large-scale datasets to perform transfer learning on tasks with smaller datasets, thereby effectively addressing the scarcity of digital holography datasets. Finally, we introduce a series of evaluation metrics to demonstrate that our model can effectively learn image features in digital holography and make rapid and accurate predictions. With proximity of number of parameters, our autofocus accuracy has improved by over 20% compared to convolutional neural networks such as Unet, with a concurrent increase in prediction speed of over 50%. Moreover, on small-scale datasets, pre-trained models have achieved an accuracy improvement of over five fold compared to direct training. These results demonstrate the potential of our transfer learning strategy in addressing the challenge of limited digital holography datasets in certain domains.