Research On Single Frame Super-resolution Algorithm Of Cell Image Based On Lens-less Microscope System And Hardware Acceleration

Lens-less microscopy systems are widely concerned by researchers in the field of cell detection because of the large imaging field and good portability.However,the system structure determines that the acquired image has a low resolution,and it is difficult to obtain detailed cell information directly.Therefore,it is of great significance to study the super-resolution algorithm of cell image based on lens-less microscope system and its hardware acceleration.Based on the analysis of the characteristics of the cell image collected by the lens-less microscope system,combined with the superiority of the convolutional neural network algorithm,a super-resolution network based on recursive convolution was built and optimized for cell image data sets with few image features and single structure.First,the network training set uses the high-definition red blood cell image collected by the optical microscope as a label,and down-samples and enlarges the size as a low-resolution training image.The network test set is a low-resolution red blood cell image collected by the lens-less microscopy system;network training taking Tensorflow as the main platform,the preliminary network model is obtained by extensively debugging the parameters such as the size of the convolution kernel,activation function and network layer,and the final network model is determined after fine-tuning each parameter according to the reconstruction effect of the test set.In order to meet the real-time calculation of the mobile platform,FPGA-based hardware acceleration is required for the designed convolutional neural network algorithm.In the first step,quantization training is performed on the premise that the reconstruction image quality loss is as small as possible,the network parameters are converted from float32 to int8,and the storage of model parameters is reduced to about 50%before quantization.The second step is to perform hardware function simulation of the super-resolution algorithm in the VCS environment.Finally,a hardware system with configurable convolutional neural network acceleration IP as the core is built in Vivado software,and after the relevant configuration of the algorithm network is completed,off-board debugging,calculation result extraction and comparative verification are performed.Super-resolves the cell images 3 times with three other classic super-resolution methods Simultaneously,subjective evaluation shows that the edge and detail texture enhancement of the cell image reconstructed by the algorithm in this paper is more obvious,and the objective image quality evaluation is improved by at least 10%.Finally,the hardware acceleration results of the FPGA platform are printed through the UART serial port.The comparison with the algorithm processing results shows that the single-frame super-resolution algorithm of the cell image can successfully achieve hardware acceleration on the FPGA platform.