Image quality is degraded in the out-of-focus region because of the depth-variant (DV) point spread function (DV-PSF) of a fluorescence microscope. Either non-blind or blind deconvolution for restoration results in limited improvement. In this work, we propose a two-step learning-based DV deconvolution (LB-DVD) to restore the out-of-focus image. In the first step, DV-PSF is predicted by a defocus level prediction convolutional neural network (DelpNet). In the second step, the extracted DV-PSF is used for DV deconvolution. To our knowledge, LB-DVD is proposed and demonstrated for the first time. DelpNet achieves an accuracy of 98.2% for predicting defocus levels of image patches ($84 \times 84$ pixels). The subsequent DV deconvolution gives rise to good performance in peak signal-to-noise ratios and structural similarity index, which are improved by up to 6.6 dB and 11%, respectively, before and after the deconvolution. As for a wide-field image, there exist different DV-PSFs within the two-dimensional fluorescence image due to the surface undulation. An overlapping weighting patch-wise LB-DVD is used in image montage to eliminate patch boundary artifacts. As a result, our LB-DVD shows the feasibility and promise to be applied to typical fluorescence microscopy in practical applications.

中文翻译：

---

使用基于学习的深度变体反卷积恢复离焦荧光显微镜图像

由于荧光显微镜的深度变化 (DV) 点扩散函数 (DV-PSF)，离焦区域的图像质量会下降。用于恢复的非盲或盲解卷积导致有限的改进。在这项工作中，我们提出了一种基于两步学习的 DV 反卷积 (LB-DVD) 来恢复失焦图像。第一步，DV-PSF 由散焦水平预测卷积神经网络 (DelpNet) 进行预测。第二步，提取的DV-PSF用于DV解卷积。据我们所知，LB-DVD 是第一次提出和演示。DelpNet 在预测图像块的散焦水平（$84 \times 84$ 像素）方面达到了 98.2% 的准确率。随后的 DV 解卷积在峰值信噪比和结构相似性指数方面产生了良好的性能，在去卷积之前和之后分别提高了 6.6 dB 和 11%。对于宽视场图像，由于表面起伏，二维荧光图像中存在不同的DV-PSF。重叠加权逐块 LB-DVD 用于图像蒙太奇以消除块边界伪影。因此，我们的 LB-DVD 显示了在实际应用中应用于典型荧光显微镜的可行性和前景。