Note: The denoising I describe in this post was not done for scientific purposes, but for artistic reasons. For proper methods on image denoising, follow the CARE paper trail.
I wanted to denoise a twenty-year-old photomicrograph. It’s one of the first scientific images I created back in 2003. It’s so ancient that I used a camera with photographic film, developed and amplified it to 15×10 cm print, and finally, scanned it to make a digital version. I used it in my undergraduate research project and find it aesthetically appealing. But, since then, it’s just been sitting on my hard drive without much use.
The image shows a multinucleated phagocytic amoebocyte on a glass slide. These are cells present in the coelomic fluid (blood) of echinoderms, They move like amoebas and engulf particles they find on their way such as debris, pathogens, and other foreign bodies. As soon as you put them on glass slides, they fuse and form giant multinucleated syncytial cells. It’s a striking behavior, but that’s for another post.
The original image is noisy. And to denoise it, I used the Noise2Void implementation in Fiji (see details and paper). It’s quite straightforward to run. You just need the single image you want to denoise for the training. Although it may take several hours to train a good model without a GPU. For this one, it took about 4h using 200 epochs with 50 steps. But the result was satisfying. The low-level graininess was gone and the image became smoother without losing detail:
Since the original dimensions weren’t large enough for what I wanted, I tried to upscale it. I had not done this recently (it used to be bad). But I found out about Real-ESRGAN, an open-source algorithm for image restoration, and gave it a try using the nice graphical interfaces from Upscaler and Upscayl. Both gave the same fine result:
Finally, I wanted to balance the colors which were off toward yellow. For that, I used GIMP‘s multitude of color tools to freely adjust them to my liking towards the grayish/blue. It’s for artistic purposes after all. But in fact, that’s closer to how they look under the microscope with proper lighting and white balance:
I’m happy with the results. Here’s the full sequence from the original to the final version: