Research On Anti-noise Segmentation Algorithm For Terahertz Inline Holographic Reconstruction Images

Terahertz light wave is widely used in all aspects of life because of its strong penetration and high biological non-destructive.Terahertz digital holography is an important part of terahertz optical imaging.However,due to the lack of sensitivity of detector,terahertz image is easily polluted by noise.And because the terahertz imaging object is small,the imaging quality will be seriously affected.Therefore,it is very important to find an anti-noise segmentation algorithm for terahertz image.In this paper,three segmentation algorithms are used for 2.52 THz digital hologram image,which are region growing,mean clustering and neural network.In order to enhance the anti-noise ability of the existing algorithms,the above algorithms are optimized to achieve better terahertz image segmentation effect.In the part of region growing algorithm segmentation,evolutionary algorithm optimization is used to segment terahertz image.Firstly,the original image is preprocessed by using bilateral filtering algorithm.Then,in order to improve the shortcomings of region growing algorithm,which needs to select seed region and growth criterion artificially,the seed region is automatically obtained by morphological corrosion.Two evolutionary algorithms,genetic algorithm and differential algorithm,are used to optimize the threshold value in the growth criterion,so as to improve the automaticity of the algorithm.Finally,we compare the convergence rate and segmentation effect of the two evolutionary algorithms,and get a more suitable evolutionary algorithm for optimizing region growth.For the research part of mean clustering segmentation algorithm,K-means clustering algorithm and fuzzy c-means clustering algorithm are used to segment one-dimensional data and two-dimensional data extracted from terahertz image respectively,and the difference between the two algorithms in segmentation effect is compared.In order to reduce the sensitivity of mean clustering algorithm to noise,a space constrained fuzzy c-means clustering algorithm is used to observe the improvement of terahertz image segmentation.Secondly,in order to reduce the over segmentation effect,the spatial position information is introduced into the objective function of spatial constrained fuzzy c-means clustering,which reduces the number of over segmentation pixels of segmentation results and enhances the filtering of clustering algorithm for noise.Finally,in the part of neural network segmentation,the terahertz image is segmented by BP neural network.By changing the noise variance of the input image of the neural network training samples,the noise residual of the segmentation results can be reduced.By adjusting the learning rate and changing the training method,the training time of the neural network is reduced.Finally,the neural network structure suitable for terahertz image is obtained.