Application of acquiring region of interest based on the YOLOv5 model in terahertz digital holography

Abstract

The imaging process of terahertz in-line digital holography is susceptible to environmental interference, and it is difficult to obtain high-quality images and image segmentation results. The challenge can be better handled by using the region of interest (ROI) condition to improve the image quality of the object region and reduce the interference of peripheral noise. In this paper, for two-dimensional 2.52 THz in-line digital holography, a method to add a variety of real backgrounds in the simulation scene is proposed, and through this method, a sufficient amount of close-to-real-scene images are produced to train the YOLOv5 model. The object ROI is extracted by the YOLOv5 model (YOLOv5-ROI). Based on the region, a new, to the best of our knowledge, support-domain-constrained phase retrieval algorithm (YOLOv5-ROI-SPRA) and an image segmentation method combined with the Otsu method (YOLOv5-ROI-Otsu) are proposed, respectively. The better results show that the expected purpose is achieved.