Cycle-PSFCNN: A Two-Stage Defocus Reasoning and Restoration Network for Microfluidic Systems

There is an issue in microfluidic systems where particles become out-of-focus during imaging due to excessively wide flow channels, which interferes with the detection of particles in the pharmaceutical manufacturing process, as well as the detection in biochemical analysis. Deep learning systems are necessary to do the restoration. However, the traditional defocus coefficient predicting networks require the aid of deconvolution operation, whose texture restoration capability is limited; while the end-to-end GAN systems introduce distortion and artifacts to the images with weak interpretability. A two-step restoration network, Cycle-PSFCNN, was proposed to combine those two tasks, forming a more robust and interpretable restoration process. The network was evaluated on both simulation and real-world datasets, and its applications in particle counting and size statistics helped microfluidic systems achieve more accurate results, reducing the statistical errors caused by defocus blur. Code is available at https://github.com/qiushu-chen/Cycle-PSFCNN.