Improving the Neural Segmentation of Blurry Serial SEM Images by Blind Deblurring

Figure 9

Evaluation results for 2D deblur-to-segmentation approach design. Left column is the blurry image and its segmentation result; SRN and DeblurGAN-V2 is the deblurring method corresponding to the study by Kupyn et al. [9], Tao et al. [8], and Tsai et al. [18], respectively; right column is our proposed deblurring methods and segmentation results. Note that each column represents a different 2D segmentation model. SRN, DeblurGAN-V2, and the proposed method are implemented through deblur-to-segmentation approaches. As shown in this figure, the deblur-to-segmentation approach performs better than a single 2D segmentation model with blur augmentation.