U-NET And Its Compound Segmentation Algorithm For Terahertz Inline Holography Reconstruction Images

In order to realize terahertz coaxial holographic reconstruction image recognition,image segmentation is usually required to extract image features.Therefore,it is necessary to study and discuss image segmentation algorithms.In this paper,the traditional threshold segmentation method based on histogram polynomial fitting and the deep learning method based on U-NET model are used to segment real terahertz images,and the two methods are combined to apply to terahertz coaxial holographic reconstruction image segmentation.This paper first introduces the characteristics of the real terahertz image of the gasket and gear to be processed.In the experiment,the real terahertz image has noise,and the pixel is divided into light and opaque parts,namely target and background,and the gasket image is The laser energy is insufficient during imaging,and the image has a background similar to the gray level of the target pixel.These problems cannot be solved by the traditional threshold segmentation method.Therefore,a composite threshold segmentation algorithm based on filtering and histogram polynomial fitting is used,that is,cropping,image expansion,NLM filtering,denoising,and histogram are performed to perform polynomial fitting to find maximum grayscale stretching,and then to histogram.The polynomial fitting is performed to find the minimum value segmentation,and the basic global threshold method and the Otsu method are compared step by step.Compared with the algorithm used by the research group Dong Ruyu,the filtering method is changed from guided filtering to NLM filtering,and the polynomial fitting order also changes,and the obtained segmentation result is better.In order to apply the existing deep learning segmentation algorithm to the continuous 2.52 THz coaxial holographic reconstruction image segmentation,U-NET can accurately segment the electron microscope neuron image under the premise of the small number of training samples,which is in line with the actual experiment in this paper.The terahertz coaxial digital holography reproduces the demand for small samples.Therefore,the influence of training set,loss function and learning rate on the results of U-NET segmentation of real terahertz images is studied,and the optimal network model is obtained.The experimental results show that the optimized U-NET model has a certain ability to distinguish the target and background of the image,and has a certain denoising ability,which can remove the noise in the image.In the experiment,the noise of the gear image is basically removed and the target is better preserved.The segmentation result is better than the traditional composite threshold segmentation method in Chapter 2,but the pad image still needs to be combined with other algorithms.Because the pre-processing steps before the polynomial fitting in the composite threshold segmentation algorithm can solve the problem of surrounding background and severe noise,the composite threshold segmentation algorithm is combined with the U-NET algorithm for real terahertz image segmentation.That is,the image is first cropped,mirrored,and NLM filtered,and then segmented by U-NET algorithm,and the experimental superposition segmentation result is repeated 10 times to eliminate the influence of randomness.Because the real terahertz image data is less,the composite U-NET algorithm segmentation of the simulated image is studied.