Application of YOLO and R-CNN Methods for Adapted Selection of Illumination Patterns in the Ghost Polarimetry

The restricted sampling efficiency of ghost imaging has impeded its extensive application to remote sensing and object positioning system. In accordance with this, our work proposed using convolutional neural networks such as the YOLOv8 and models from the R-CNN family to select segments on the ghost image plane with the subsequent formation of a set of modified illumination patterns. This set differs from the original one in the presence of areas of high spatial resolution in the found segments. Thus, the use of deep learning makes it possible to implement adaptive selection of illumination patterns to increase the quality of reconstructed ghost images. In addition, an auxiliary polarization channel was also introduced into the imaging system, which can be used to speed up data acquisition or to recover the polarization pictures of objects.