Research on training set in image segmentation of terahertz digital holographic reconstructed image based on convolutional neural network

In the case of a certain wavelength, digital holographic imaging is larger than other imaging methods can obtain finer target structures. Therefore, terahertz digital holography technology has received more and more attention. Among them, the image segmentation of reproduced images has important application value. The training set is very important in image segmentation based on convolutional neural network; and currently there are not enough terahertz digital holographic images, and there is no standard training set for real images. For this reason, this article self-made a 2.52THz simulation reconstruction image based on angular spectrum and phase retrieval algorithm as a training set, the image size of which is 256×256. The test uses a real 124×124 terahertz reconstructed image, and expands to 256×256; and objectively evaluates the segmentation results at this image resolution. Compared with other training sets, the results show that it can be better segmented by using the data set established in this paper.