February_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 1 14 of the use of these methods are described. In BPFA meth- odology there are anumber of tunableparameters that are used to increase the efficiency of the algorithms to recon- struct the images. Again, this article cannot go through all the details of the processes involved, but will simply show the results and refer the reader to other publications that discuss the methodology involved. [4-8] The first example shows the reconstruction of the famous “Barbara” image. Figure 3 shows the original image sub-sampled to 25%and reconstructedusingBPFA. The two example reconstructions highlight the “tuning” of the reconstruction parameters that can be achieved and the quality of the retrieved image. The example shown in Fig. 3 involves an image that has been acquired fully sampled and then sub-sampled for demonstration purposes. The final example shows an image reconstructed from an experiment where the image was acquired as a sub-sampled image and then reconstructed. Figure 4 shows a traditionally acquired atomic resolution Z-contrast STEM image of Ceria, a 6.25% sub-sampled line-hop image of the same sample in the same position (beam damage is not an issue at these doses for this sample), and the reconstructionperformedon theaforementioned sub-sampled image. Todetermine theaccuracyof the reconstruction, Fig. 4c is compared to Fig. 4a by twometrics, peak signal-to-noise ratio (PSNR) and cross correlation. Figure 4c has a PSNR of 20.6752 dB and a maximum cross correlation of 0.75037 when compared to Fig. 4a, both of which fall withinacceptable limits for interpreting the image directly. This approach increases the speed of the acquisition by 16x while reducing the overall dose by the same amount, and yet the quality of the image is essentially the same. CONCLUSION The results discussed in this article demon- strate that it is possible to determine an optimal approach to forming the most efficient image in any scanned imaging system. By moving beyond hardware-only defined solutions for imaging beam sensitive samples and employ- ing compressive sensing/inpainting/machine learningmethods to reconstruct images that are sub-sampled, the spatial and temporal resolu- tion of images can be increased. This approach opens up a wide range of materials and dynamic Fig. 3 (a) Public domain test image, Barbara. (b) 25% randomly sub- sampled image of Barbara. (c) Reconstruction of (b) with deliberately poor reconstruction parameters. (d) Reconstruction of (b) with optimized reconstruction parameters. Fig. 4 (a) 512 x 512 fully sampled atomic resolution Z-contrast image of Ceria. (b) 6.25% sub-sampled line-hop image acquired with the same beam conditions at the same location. (c) Reconstruction of the 6.25% sub-sampled image using BPFA. (a) (b) (c) (d) (a) (b) (c)