Research On Low-resolution Image Fusion Algorithm Based On Deep Learning

The lensless microscopic imaging system provides a good solution for the development of cell-portable detection equipment due to its low cost,miniaturization and other advantages.However,the cell images collected by this system have problems such as low resolution and less feature information.In order to facilitate the observation and detection of cell morphology,this paper mainly studies the low-resolution cell image super-resolution reconstruction and image feature fusion algorithms suitable for lensless imaging systems.In view of the above problems,it is difficult to effectively solve the above problems by using traditional image processing algorithms.This paper proposes image super-resolution algorithms based on generative confrontation networks and image feature fusion algorithms based on convolutional self-encoding networks.Completed super-resolution reconstruction and feature fusion of lensless low-resolution cell images.First,the method of preprocessing and data set establishment of blood cell samples for the lensless system is studied,and the established data set is used for super-resolution reconstruction and feature fusion network training test.Secondly,the generative confrontation super-resolution network(LSRGAN,Lensless Super-resolution Generative Adversarial Network)is optimized and improved in terms of network structure,loss function,training strategy,etc.After the network model is stabilized,it is tested and evaluated by image quality evaluation indicators.The reconstruction results show that:comparing the super-resolution algorithm LSRGAN in this paper with reconstruction algorithms such as Bicbic and SRGAN,the algorithm proposed in this paper performs well in both information entropy and grayscale gradient,and the subjectively visible cell edge and texture information is significantly enhanced.After super-resolution reconstruction of the cell image,the convolutional auto-encoding fusion network is used to fuse the cell characteristics of the same type of tissue.The fusion result is a collection of "virtual cells",which represents the statistical results of the characteristic morphology of similar tissues.Multiple groups of normal and abnormal cell images(8 in each group)were mixed and randomly fused in different proportions for testing,and the results were analyzed and compared using non-reference evaluation indicators such as information entropy,gray gradient,and mean square error.Then use the fused images to establish a retrieval database combined with the traditional KNN classification algorithm to retrieve and classify unknown cells.The results of the fusion show that the fusion algorithm in this paper improves the information entropy by 10.37%compared with the traditional fusion algorithm.The standard deviation and gray gradient are significantly higher than the traditional algorithm,only the gray gradient is slightly lower than the neural network algorithm PCNN,which proves that the fusion network can enrich the image feature details features and enrich the texture information.The image retrieval results provide a reference basis for the preliminary auxiliary medical diagnosis.