Multi Frame Pathological Image Super Resolution Reconstruction Based On Deep Learning

Cytopathological analysis is of great significance for the examination and diagnosis of diseases,but the vision scope is small and the acquisition speed is slow under the high power objective.Super-resolution reconstruction(SR)technology can reconstruct pathological microscopic images by up sampling and obtain high resolution images,which greatly improves the efficiency.The super-resolution reconstruction methods based on deep learning hava achieved good results.However the general methods use artificial downsampling images as low resolution images of training set.When the images have unknown degradation,the artificial downsampling images can not represent the real images,so the methods have limited effect.Moreover,the existing reconstruction methods are difficult to take into account both the high frequency similarity and the low frequency similarity.In order to solve the above problems,this thesis studies the super-resolution reconstruction technology of pathological microscopic image based on deep learning.The main work is as fallows:(1)A cell data set including groups of high and low resolution image pairs was collected by pathological scanning system,and a set of data preprocessing process was designed.(2)An information fusion and super-resolution image pre-generation model is constructed.The network fuses the information of multiple low-resolution images and generates a super-resolution image which has high low-frequency similarity with the reference image.The network uses iterative fusion for information fusion,then uses multiple residual blocks for feature mapping,finally uses sub-pixel convolution for reconstruction.(3)A super-resolution image texture generation model is constructed.On the basis of ensuring the low-frequency similarity between the generated image and the original image,the network adds more realistic texture details to the generated image.The network is based on the structure of generative adversarial networks(GAN).The generator first fuses the information,then uses the encoding and decoding structure for feature extraction and image reconstruction.Multiple residual blocks are used for feature mapping.The discriminator uses Markov GAN mechanism.The gradient weighted loss is designed to make the network pay more attention to generate image texture.Finally,the results are compared with several classical SR methods.The experimental results show that the information fusion and pre generation model can generate images which have high low frequency similarity with the original image,and the texture generation model can generate images with good performance in both subjective and objective indicators.