4/24/24. Semantic Segmentation (Fancy image editing)
Frank L -
Remember that ink project that I mentioned a month or so ago? Well, I’m back on it now. Like I said in my last post, the GAN project that I’ve assisting is slowly coming to an end, so there isn’t much left me to do anymore on that. As a result, I’ve went back to helping on another main project, the Aerosol Jet Printing one with the cool lines of ink.
The task I’m working on now is called semantic segmentation, which is basically just a fancy way of editing an image to help computer software better process visual information. Semantic segmentation is method of classifying and assigning labels to specific pixels within in an image. This helps deep learning models efficiently analyze the shapes and forms of objects and figure out where the important parts are. In order to do this, I have to segmentation masks, which are portions of the image that are differentiated from everything else.
A really simple way to do semantic segmentation is by just painting over the important parts of an image with a labeling software. This technique is sometimes pretty inefficient, so another way is by simply drawing in colored polygons over the image. That’s what I did in the image below. I put both the original image and the segmentation map version of it below. The red part is labeled as the central line, the green parts are the ink overspray, and the blue specks are the solvent overspray.