Research on Single-Frame Super-Resolution Reconstruction Algorithm for Low Resolution Cell Images Based on Convolutional Neural Network

Abstract

Aiming at the problem of low resolution and low amount of contented of cell images collected by lens-less cell detection systems consisting of CMOS image sensors and microfluidic channels, a novel CSRNet (Cell Super Resolution Network) reconstruction network for lens-less cell detection systems is proposed in this paper. First, the images of cells on blood smears were collected with a microscope (MshOt model), and using a threshold segmentation algorithm to divide it to 80×80 high-resolution (HR) reference image for training. After segmentation, the HR cell image is downsampled by the bicubic downsampling algorithm to obtain a 20×20 low resolution (LR) cell image. Training the CSRNet reconstruction network with a training set consisting of HR and LR cell images. Then the cell image which was collected by the lens-less cell detection systems is segmented. The segmented cell image is inputed into a trained network model, and a high-resolution cell image with more detailed information is as a output. The experimental results show that the effect of the proposed CSRNet network reconstruction is better than both the bicubic interpolation reconstruction and the FSRCNN reconstruction in terms of subjective visual and objective evaluation indicators. Therefore, the CSRNet network can be used to improve the image resolution which collected by lens-less cell detection systems.