Research On Super-resolution Algorithm Of Leukocyte Image

The morphology of leukocytes varies greatly,and the accuracy of the classification of leukocytes in the blood analysis system is closely related to its image quality.In order to ensure that the white blood cell image clear,this paper designs a white blood cell image super-resolution reconstruction algorithm based on generating an adversarial network.On the premise of ensuring that the detailed texture of the white blood cell image after over-score is clear,its classification accuracy is consistent with the original image.Main tasks as following:1.White cell image super-resolution algorithm based on the generative adversarial network.First,the high-resolution pictures are down-sampled to simulate the image degradation process,and then the paired high-low resolution white blood cell images are processed into the LMDB format to speed up the reading speed of the pictures on the network.Then,based on the super-resolution research of the original generated adversarial network,improvements are made,the nested dense residual blocks are used to replace the basic residual blocks in the original network structure,and the training method is optimized by combining perceptual loss,adversarial loss and L2 loss.,So that the network can learn clear texture details.Finally,cross-validate the improved algorithm in this paper to ensure the stability of the algorithm model.2.Design of WBC classification network based on deep learning.This paper designs a classification network based on deep learning,and optimizes the number of layers and width of the network to make the classification more accurate.Then,the optimized network is used to classify the images before and after the super-score algorithm processing.The results show that the classification accuracy of the super-score image is close to the original image.3.Experimental design.This article collected three different white blood cell public data sets from the Internet,and selected a total of 863 high-quality white blood cell images from it,using the down-sampling method to obtain the corresponding low-resolution images,as a data set for training super-resolution algorithms.At the same time,the data set was amplified by different degrees of translation and rotation.Finally,31,380 white blood cell images were obtained to meet the classification network's demand for white blood cell data.In addition,the design experiment processed the improved super-resolution algorithm with three interpolation methods and four learning-based super-resolution methods under the same conditions to process the white blood cell image,and compared the results subjectively and objectively.The results show that the image processing effect of the improved algorithm is better.This topic uses a super-resolution algorithm based on generating an adversarial network to perform super-resolution processing on white blood cell images.The PSNR and SSIM of the processed images are 35.798 and 0.9483,respectively.In addition,the classification experiment of white blood cell images before and after over-score was performed.From the classification results,the improved super-resolution algorithm processed white blood cell images to ensure that the classification accuracy rate was close to the original image,which was 95.45%.