Image Segmentation, UNet, and Deep Supervision Loss Using Keras Model

Image segmentation entails partitioning image pixels into different classes. Some applications include identifying tumour regions in medical images, separating land and water areas in drone images, etc. Unlike classification, where CNNs output a class probability score vector, segmentation requires CNNs to output an image. Accordingly, traditional CNN architectures are tweaked to yield the desired result. An array of architectures, including transformers, are available to segment images.