面向显微视觉的端对端去模糊模型  

End-to-end deblurring model for microscopic vision

作　　者：徐征[1] 何佳珩 王彦琪 王晓东[1] 任同群[1] XU Zheng;HE Jiaheng;WANG Yanqi;WANG Xiaodong;REN Tongqun(College of Mechanical Engineering,Dalian University of Technology,Dalian 116081,China)

机构地区：[1]大连理工大学机械工程学院,辽宁大连116081

出　　处：《光学精密工程》2024年第20期3047-3058,共12页Optics and Precision Engineering

基　　金：国防基础科研计划资助项目(No.JCKY2022203B006);中央高校基本科研业务费资助项目(No.DUT24LAB112)。

摘　　要：显微视觉测量在微装配等领域中应用广泛,受成像景深等因素的影响,图像会出现多重离焦模糊现象,影响后续准确测量,而显微自动对焦技术虽然可以缓解离焦问题,但耗时较长,难以适应高效生产要求。本文提出了将模糊度判别和多分支恢复相结合的端对端去模糊模型,建立了分块、判别、去模糊、融合的分而治之策略:首先将一幅图像切割成子图像组,同时送入判别器和恢复网络;在判别器中,通过傅里叶变换等获取频域分布,再利用Vision-Transformer网络从频域图中提取具有全局关联性的频域深层模糊特征,然后对模糊度进行判别输出。根据判别结果,由多分支恢复网络对不同模糊度的子图像进行定向恢复,最后融合拼接处理后的子图像,获得高清晰度的图像。实验结果表明,本文提出的模型能有效恢复多重模糊的显微图像,判别准确率达0.94,而模糊图像经过多分支恢复网络处理后,PSNR指标平均提升了6.3。The measurement of microscopic vision is commonly used in micro-assembly and other fields.However,due to limitations such as depth of field in microscopic imaging,the image may appear blurred and affect the accuracy of measurement.Although the technology of auto-focusing in optical microscopy can alleviate defocusing problems,it will be too time-consuming to adapt to the requirements of efficient production.Herein,an end-to-end deblurring model that integrates blurring discrimination and multibranch recovery was presented,in which a divide-and-conquer strategy of chunking,discrimination,deblurring,and fusion was established.Firstly,the image was divided into sub-images,which were then simultaneously processed by a discriminator and a recovery network.The discriminator employed the Fourier transform to obtain the frequency-domain map of the sub-images.From the frequency domain map,the Vision Transformer network extracted deep blur features with global correlation.The output of the blurring degree was then discriminated.The multi-branch recovery network was used to directionally recover sub-images with different blurring degrees based on the discriminative output.Finally,the spliced sub-images were fused to obtain high-resolution images.The experimental results indicate that the model can effectively restore multi-blurred microscopic images,with a discriminator accuracy reaching 0.94.Moreover,after undergoing processing by the multi-branch restoration network,the PSNR metric shows an average improvement of 6.3.

关 键 词：微装配 显微视觉 模糊度判别 频域处理 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]