Research On Microscopic Image Reconstruction And Enhancement Algorithm Based On Deep Learning

Photonic crystal material as a material with special optical properties and cell compatibility is often used in organ chips.Usually,it is used as substrate for cell culture.It can indirectly observe the mechanical properties and mechanical properties of cells.The existing analysis methods for photonic crystal films mainly use imaging equipment for image acquisition,and then indirectly analyze the response of cultured cells to external stimuli through image processing methods.However,when imaging photonic crystal samples,several problems often appear,including inaccurate image focus,poor image quality(low resolution,low signal-to-noise ratio),and difficulty in three-dimensional reconstruction of the surface.This article mainly uses deep learning-based algorithms to study the above three problems to realize the reconstruction and enhancement of optical microscopy images.The specific research work is as follows:1)This article uses the transfer learning strategy to realize the autofocus task and patch detection of the diseased area.In this study,the auto-focus task is transformed into a multi-classification task,and the auto-focus is realized on the Yeast Z-Stacks,MCF-10 A,and PCF datasets by using the pre-trained Res Net model.For the patch detection task of the diseased area,this research combines the VGG-19 network of transfer learning and the probability-weighted voting algorithm,selects the appropriate patch size,and processes each patch separately to achieve virus infection in a single-hole field of view proportion calculation.2)This paper proposes a DN-SRNet for joint denoising and super-resolution reconstruction tasks to enhance the image of photonic crystal film samples.The model is composed of multi-scale residual-dense attention groups,and a unique loss function is designed for this task.The results on the W2 S and FMD datasets show that the algorithm can obtain better PSNR and SSIM scores than the baseline.3)Aiming at the problem of photonic crystal film surface reconstruction,this paper uses Gaussian mix-ture model to generate a simulated photonic crystal film height dataset.The corresponding relationship between the light intensity of the collected image and the tilt angle of the sample is measured on real samples,and use the rasterized optical analysis method to generate the corresponding illumination map.Then,this paper trains a conditional generative adversarial network on this dataset,realized the reconstruction of the surface height map of the corresponding sample from the input illumination map,thus the three-dimensional reconstruction of the upper surface of the photonic crystal film is finished.