Research On Segmentation Algorithm Of Terahertz Inline Holographic Reconstructed Images

Terahertz image has a wide range of application value,but due to the current problems of large imaging noise and small image size,it is difficult to extract the features of terahertz image,so it is necessary to study its image segmentation algorithm.This thesis first proposes a Markov Random Field Image Segmentation(MS-MRF)algorithm based on the mean shift clustering to determine the number of clusters according to the characteristics of such images.The number of clusters is automatically determined through the gray histogram of the image,and the number of clusters is determined by the gray histogram of the image.The real terahertz image is segmented.This paper also uses the U-Net model to train a convolutional neural network suitable for segmentation of terahertz images,and deepens and improves the network structure based on the U-Net network to obtain a better segmentation effect of the terahertz coaxial holographic image.This paper first analyzes the characteristics of five real terahertz images,and uses threshold segmentation to process the images.Because the laser energy is low,the target size is small,the background noise of some real terahertz images overlaps with the target image,and the gray level is similar.Therefore,it is difficult to segment the image using a single traditional threshold segmentation algorithm.Therefore,this paper proposes the MS-MRF algorithm,which uses the mean shift filter to perform preliminary clustering of the image,and then uses the gray-level histogram-based peak positioning algorithm to automatically determine the number of clusters,and finally uses Markov random field to achieve automatic image segmentation.A better segmentation effect.This paper also studies the image segmentation algorithm of convolutional neural network based on U-Net.Only the normal image segmentation data set is not ideal for the segmentation of gaskets and G-shaped images.In order to train it as a network suitable for segmenting the terahertz coaxial holographic image,this paper establishes the same feature as the terahertz image.The noisy terahertz simulation image is used as a data set,and multiple hyperparameters are adjusted to obtain the optimal network model.At the end of this paper,an image segmentation algorithm with improved U-Net convolutional neural network structure is studied.In order to obtain better training effects and enhance network performance,this paper analyzes the segmentation effects of 22,28,and 34-layer convolutional neural networks.At the same time,a residual connection structure is added to the network to solve the gradient dispersion and gradient explosion caused by the excessive number of network layers.In order to improve the training effect,a dropout layer is added to the network.The final segmentation effect is better than the MS-MRF algorithm and the original U-Net network,and overall it has the best segmentation effect.