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

Figure 6

Evaluation results from the M1 dataset. The first to the fourth column represents the 5% training dataset, 10% training dataset, 20% training dataset, and 50% training dataset, respectively. The first and last rows are blurry images and clear images. From the second row to the fifth row represents the deblurring method corresponding to the study by Nah et al. [5], Kupyn et al. [9], Tao et al. [8], Tsai et al. [18], and our proposed method. As shown in this figure, our proposed method can reconstruct myelin sheaths and intracellular features with accuracy.