Research On Key Technologies Based On Microscopic Global Automatic Imaging

Microscopes play an indispensable role in the field of microscopic inspection and are widely used in industrial defect detection,biomedical research,and scientific research.Because of its high magnification,the imaging has the characteristics of a shallow depth of field and a small field of view and cannot directly complete the imaging observation of the whole area of the test sample.Manual focusing of the sample is required,and the sample position needs to be manually adjusted to achieve the observation of different regional features.Therefore,it is of great importance to study microscopic full-field automatic imaging.The key technology of microscopic full-field automatic imaging lies in the automatic focusing and image stitching of microscopic images.The automatic focusing technology aims to accurately complete the acquisition of clear images,while the image stitching technology reconstructs a full-field image from a large number of small field-of-view images.The technical difficulties mainly include:(1)how to improve the efficiency of autofocus under the condition of ensuring the focusing accuracy;(2)how to image clearly the samples with surface concavity greater than the depth of field to provide stable features for image stitching;(3)how to develop stitching strategies to avoid the accumulation of errors for large-scale microscopic images.To address the issues raised above,the text primarily accomplished the following work elements:(1)The hill-climbing algorithm combining coarse and fine steps is improved for the existing focus search strategy.Based on the different characteristics,the existing focus evaluation functions were studied,and different focus evaluation functions were tested by a series of images acquired by the microscope with equal interval displacement in Z-axis,and the performance characteristics of different evaluation functions were analyzed,and a variable-step intelligent search focusing strategy with multiple evaluation functions was designed.The whole focusing process adopts the variable step length technique to achieve fast focusing on the detection object by shortening the search step.(2)Bayesian convolutional neural networks are proposed for the focus prediction of out-offocus images.More and more deep learning methods are introduced for focus prediction of singleframe images,yet almost all network models overly trust their outputs and do not give any warning even if they output wrong results in the face of unknown samples.In this paper,we propose a Bayesian convolutional neural network that can model the uncertainty of the prediction results based on the prediction of the focal position of the out-of-focus system.And a set of screening mechanisms is established,which can eliminate some of the wrong predictions by setting the threshold of uncertainty.The experimental results show that the network model proposed in this paper can output higher levels of uncertainty on unknown samples compared with the same type of samples,and the established screening mechanism can effectively reduce the prediction error of the model on unknown samples.The final error ranges of 0.37 ± 0.46 μm and 0.83 ± 1.17 μm for the two samples in the public data set were better than the pre-screening 0.40 ± 0.66 μm and 1.08 ± 1.78 μm.(3)The large-scale microscopic image stitching task is realized,and the cumulative errors generated during the stitching process can be eliminated.The article first analyzes the imaging characteristics of planar and non-planar objects and then investigates the depth-of-field expansion technique using multi-focus fusion for the problem of an unclear global field of view caused by the degree of concavity of non-planar objects larger than the depth-of-field range.Clear images can provide stable features for image alignment,and this paper uses the SIFT algorithm for alignment.An optimization method of the global registration process is realized for large-scale microscopic stitching.The overdetermined equations are constructed by using the registration relationship of adjacent images,and the least squares method is used to solve the equations.Finally,the absolute coordinates of each image can be obtained to realize the stitching process,and the stitching results are not affected by the cumulative error.(4)Eventually all algorithms were systematically integrated,and the effectiveness of the system built in this paper was demonstrated by scanning tests on tissue sections with planar features versus pharmaceutical powders with non-planar features.The microscopic full-field automatic imaging technique studied in this paper achieves automatic focusing and large-scale image stitching of the inspection field of view,and the obtained research results can be used for the automatic modification of manual microscopes to realize the automatic imaging task of microscopic full-field.